Announce: dumpfs v0.01 - common RAS output API

Announcing dumpfs - a common API for all the RAS code that wants to save data during a kernel failure and to extract that RAS data on the next boot. The documentation file is appended to this mail.

ftp://oss.sgi.com/projects/kdb/download/dumpfs - current version is v0.01, patch against 2.6.8-rc2.

This is a work in progress, the code is not complete and is subject to change without notice.

dumpfs-v0.01 handles mounting the dumpfs partitions, including reliable sharing with swap partitions and clearing the dumpfs partitions. I am working on the code that reads and writes dumpfs data from kernel space, it is incomplete and has not been tested yet. After dumpfs_kernel is working, dumpfs_user is trivial. The code is proof of concept, some sections of the API (including polled I/O and data compression) are not supported yet, and some of the code is ugly.

Why announce incomplete and untested code? Mainly because RAS and kernel dumping are being discussed at OLS this week. Since I cannot be at OLS, this is the next best thing. Also the dumpfs API has stabilized for the first cut, so it is time to get more discussion on the API and to determine if it is worth continuing with the dumpfs approach. If dumpfs is discussed at OLS then I would appreciate any feedback.

Questions for the other people who care about RAS (which rules out most of the kernel developers) -

* Is using a common dump API the right thing to do?

Obviously I think that this makes sense. At the moment every bit of RAS code has its own dedicated I/O mechanism, not to mention its own user space tools to interface with the kernel, and to initialize, extract and clear its own data.

dumpfs consolidates a lot of common code that is scattered over several RAS tools. dumpfs removes the need for special RAS tools to extract dump data on reboot, instead standard user space commands will do the job.

* Is overloading mount the best approach?

Making mount dumpfs share the partition with swap is ugly. OTOH most of the existing code that dumpfs is intended to replace makes no attempt to verify its partition usage. At least dumpfs tries to verify its partition data, ugly though the code is.

* Does the dumpfs API need to be extended or even replaced, either in kernel or in user space?

One obvious extension is to make compression selective, so that some sections of the file can be compressed and others be in clear text. The lcrash header springs to mind. Omitted for now since this version does not support compression yet.

* How do we get a clean API to do polling mode I/O to disk?

One thing that is absolutely required for reliable RAS output is a polling mode method. netdump is available for the network, we need the equivalent for disk I/O. What is the best way to integrate polling mode I/O into the block device subsystem?

If the people who care about RAS think that a common RAS output API is worthwhile then I will continue working on dumpfs. Otherwise it will be just another idea that did not get taken up, and each RAS tool will continue to be developed and maintained in isolation.

==== 2.6.8-rc2/Documentation/filesystems/dumpfs.txt ====

dumpfs provides a common API for RAS components that need to dump kernel data during a problem. The dumped data is expected to be copied and cleared on the next successful boot.

dumpfs consists of two layers, with completely different semantics. These are dumpfs (kernel only) and dumpfs_user (user space view of any saved dump data).

dumpfs uses one mount for each dump partition. Each dumpfs partition can be mounted with option share or noshare, the default is noshare. The only allowable user space operations on a dumpfs partition are mount and umount, user space cannot directly access the dumpfs data. Each dumpfs partition is mounted with "mount -t dumpfs /dev/partition /mnt/dumpfs". /mnt/dumpfs must be a directory; it never contains anything useful but the mount semantics require a directory here.

A shared dumpfs partition will normally coexist with a swap partition; the dumpfs superblock is stored at an offset which leaves the swap signature alone. A shared dump partition has no superblock on disk until the first dump file is created. Mounting a dumpfs partition with "-o clear" will completely zero the dumpfs superblock, including the magic field. This ensures that old dumpfs data in a shared partition will not be used, its contents are unreliable because of the data sharing.

When mounting a shared dumpfs partition, no check is made to see if the disk contains a dumpfs superblock. Mounting a dumpfs partition with -o share will only share with a swap partition, it will not share with any other mounted partition.

A non-shared dumpfs partition must have a superblock before being mounted. mkfs.dumpfs and fsck.dumpfs (only used for non-shared partitions) are trivial. Mounting dumpfs with "-o noshare,clear" will clear the metadata in the dumpfs superblock, but preserve the magic field.

mkfs.dumpfs

#!/bin/sh
dd if=/dev/zero of="$1" bs=64k count=1
echo 'dum0' | dd of="$1" bs=64k seek=1 conv=sync

fsck.dumpfs

#!/bin/sh
true

Each dumpfs partition can be mounted with option poll or nopoll, the default is poll. Poll uses low level polled mode I/O direct to the partition, completely bypassing the normal interrupt driven code. This is done in an attempt to get the data out to disk even when the kernel is so badly broken that interrupts are not working. Poll requires that the device driver for the dumpfs partition supports polling mode I/O. Nopoll uses the standard kernel I/O mechanisms, so it is not guaranteed to work when the kernel is crashing. Nopoll should only be used when your device driver does not support polling mode I/O yet; you must accept that dumpfs may hang waiting for the I/O to be serviced.

Another option when mounting a dumpfs partition is to specify the size of its data buffer, in kibibytes. This buffer is permanently allocated as long as the dumpfs partition is mounted, it is only used when writing RAS data via dumpfs. The buffer size will be rounded up to a multiple of the kernel page size. The default is buffer=128.

The user space view of the RAS data held in the dumpfs partitions is created by "mount -t dumpfs_user none /mnt/dumpfs". It logically merges and validates all the dumpfs partitions that have been mounted and provides a user space view of the files that have been written to dumpfs. The only user space operations supported on dumpfs_user are llseek, read, readdir, open (read only), close and unlink. Just enough to copy the files out of dumpfs_user and remove them. User space cannot write to dumpfs_user.

The kernel can write to files held in dumpfs partitions, to save RAS data over a reboot. Note that when kernel RAS components write to dumpfs they do _not_ use the normal VFS layer, it may not be working during a failure. Instead a RAS component makes direct calls to the following dumpfs_kernel functions.

dumpfs_kernel_open("prefix", flags)

Create and open for writing a file in dumpfs. It returns a file descriptor within dumpfs.

The dumpfs filename is constructed from "prefix-" followed by the value of xtime in the format CCYY-MM-DD-hh:mm:ss.n, where n starts at 0 and is incremented for each dumpfs file in the current boot.

There is no requirement that a dumpfs_user mount point exist before the kernel can dump its data. The first call to dumpfs_kernel_open will automatically create a kernel view that merges all the mounted dumpfs partitions. The first call to dumpfs_kernel_open also writes the dumpfs superblocks to any shared partitions.

Flags select compression, if any.

dumpfs_kernel_open() is the simple interface. It automatically stripes the data across all dumpfs partitions that are not currently being used.

Most RAS code will open one dump file at a time, mainly because most users will only have one dumpfs partition. The dumpfs code has a module_parm called dumpfs_max_open, with a default value of 1.

dumpfs_kernel_bdev_list() dumpfs_kernel_open_choose("prefix", flags, bdev_list)

Some platforms may need to have multiple output streams open in parallel. For example a system with large amounts of memory and multiple disks may wish to assign different sections of memory to each cpu and to write to separate partitions.

dumpfs_kernel_bdev_list() returns the list of usable dumpfs partitions. If all partitions are in use then the list is empty.

dumpfs_kernel_open_choose() opens a file using only the selected bdev entries.

Systems that use concurrent parallel dumps should set module_parm dumpfs_max_open to a suitable value.

Note: The following problems are inherently architecture and platform specific and are outside the scope of dumpfs. That is not to say that we should not have an API for handling these problems on large systems, but it would be a separate API from dumpfs.

Deciding which cpus to use for parallel dumping.
Deciding which block devices each cpu should use.
Getting the chosen cpus into the RAS code.
Assigning the range of work to each cpu and each partition. Watching the dumping cpus for problems, recovering from those problems
and reassigning the work to another cpu.
Reconstructing the parallel dumps into a format for analysis. dumpfs_user
makes each dump file available to user space, but some code may be required to merge the separate files together.

dumpfs_kernel_close(fd)

Sync the file's data to disk, close the file and update the dumpfs metadata.

dumpfs_kernel_write(fd, buffer, length)

Write the buffer at the current dumpfs file location. The data may or may not be written to disk immediately. It returns the current location, including the data that was just written.

For performance, the dumpfs data is striped over all the assigned partitions, in round robin. The stripe unit is the minimum of the buffer= value across all the assigned partitions.

dumpfs_kernel_read(fd, buffer, length)

Read the buffer from the current dumpfs file location. It returns the current location, including the data that was just read.

dumpfs_kernel_llseek(fd, position)

Set the current dumpfs file location. It returns the previous location. Only absolute seeking is supported.

dumpfs_kernel_sync(fd)

Sync the file's data to disk and update the dumpfs metadata.

dumpfs_kernel_dirty_shared()

Returns true if any shared partitions have been dirtied, in which case the kernel must be rebooted after all the RAS components have completed their work.

dumpfs_kernel_all_polled()

Returns true if all dumpfs partitions can support polling mode I/O. Otherwise the RAS code that calls dumpfs should enable interrupts, if at all possible.

Sample /etc/fstab entries for dumpfs partitions.

  /dev/sda2  /mnt/dumpfs  dumpfs  defaults  0 0
  /dev/sdb2  /mnt/dumpfs  dumpfs  share     0 0
  /dev/sdc7  /mnt/dumpfs  dumpfs  nopoll    0 0

Sample code in /etc/rc.sysinit to save dump data from the previous boot. If you are sharing dumpfs with swap, these commands must be executed before mounting swap. Note that dumpfs does not require any special user space tools to poke inside partitions to see if there is any useful data to save, everything is a file.

  # mount all the dumpfs partitions
  mount -a -t dumpfs
  # merge all dumpfs into dumpfs_user on /mnt/dump
  mount -t dumpfs_user none /mnt/dump
  # copy the data out
  (cd /mnt/dump; for f in `find -type f`; do echo saving $f; mv $f /var/log/dump; done)
  # drop dumpfs_user
  umount /mnt/dump
  # clear all the dumpfs metadata
  umount -a -t dumpfs
  mount -a -t dumpfs -o clear
  umount -a -t dumpfs

rc.sysinit will later mount the swap partitions, then mount all the other partition types. That will remount the dumpfs partitions, ready for the next kernel crash.

We hope to not have to. The current plan is to use kexec: at boot time, do a kexec preload of a small (16MB) kernel image. When the main kernel crashes or panics, jump to the kexec kernel. The kexec kernel will hold a new device driver for /dev/hmem through which applications running under the kexec'ed kernel can access the crashed kernel's memory.

Write the contents of /dev/hmem to stable storage using whatever device drivers are in the kexeced kernel, then reboot into a real kernel again.

That's all pretty simple to do, and the quality of the platform's crash dump feature will depend only upon the quality of the platform's kexec support.

People have bits and pieces of this already - I'd hope to see candidate patches within a few weeks. The main participants are rddunlap, suparna and mbligh.

Hmm. I think this will require one of the kernels to run at a non-default address in physical memory.

And at this point I don't quite see why you would need /dev/hmem, as opposed to just using /dev/mem.

Or will the crashing kernel save and compress the core dump to somewhere in ram and the dump kernel read it out from there?

Which will largely depend on the quality of it's device drivers...

I'm sorry I missed you then. Unfortunately this is my busiest season at work so I wasn't able to make it to OLS this year :(

Does anyone have a proof of concept implementation? I have been able to find a little bit of time for this kind of thing lately and have just done the x86-64 port. (You can all give me a hard time about taking a year to get back to it :) I am in the process of breaking everything up into their individual change patches and doing a code review so I feel comfortable with sending the code to Andrew. So this would be a very good time for me to look at any code for reporting a crash dump with a kernel started with kexec.

Eric

This differs a little from your earlier suggestion of requiring a kernel to run from a non-default address. Martin suggested simply reserving about 16MB of area in advance, so that just before kexecing the new kernel with mem=16M, we save the first 16MB away into the reserved space. /dev/hmem (oldmem ?) is a view into the old kernel's memory, as opposed to /dev/mem.

Yes, Hari has a nice POC implementation - it might make sense for him to post it rightaway for you to take a look. Basically, in addition to hmem (oldmem), the upcoming kernel exports an ELF core view of the saved register and memory state of the previous kernel as /proc/vmcore.prev (remember your suggestion of using an ELF core file format for dump ?), so one can use cp or scp to save the core dump to disk. He has a quick demo, where he uses gdb (unmodified) to open the dump and show a stack trace of the dumping cpu.

Regards
Suparna

Suparna Bhattacharya ([email protected])
Linux Technology Center
IBM Software Lab, India

If you are prepared to say "dump device is IDE" the driver ends up pretty tiny - maybe 4K for PIO mode.

--Alan Cox <[email protected]> wrote (on Wednesday, July 28, 2004 11:46:06 +0100):

After kexec, we shouldn't need such things, do we? Before it, Linus won't take the patch, as he said he doesn't like systems in unstable states doing crashdumps to disk ...

M.

And what does kexec do.. it accesses the disk. A SHA signed standalone dumper is as safe as anything else if not safer.

Alan

--Alan Cox <[email protected]> wrote (on Wednesday, July 28, 2004 15:06:27 +0100):

But it's reading, not writing ... personally I'm happier with that bit ;-)

M.

And it is only reading to preload the dumper in memory. This happens before the system crashes.

All that happens at crash dump time is we hand off control to the dumper. kexec is just the mechanism to switch from the kernel to the dumper.

It is attractive to make the dumper based on the current linux kernel but that is by no means a requirement.

Eric

I was using mcore a few years back and it didn't need to read anything to launch the new kernel image with bootimg, the new kernel image was stored in a safe place in memory IIRC

Thanks having a signature or a checksum on the dumper stored in memory is one piece of this puzzle that I keep forgetting. Checking the signature before we trust the code is important.

Eric

That sounds about right. I have a read-only version of Etherboot that pretty much does that.

A large challenge is how do you get the IDE device into a sane state before you use it to dump core. How do you stop on-going DMA transactions. What about implementation errata etc. Do you need to reprogram the drive to work in a PIO mode again. I can't quite remember by I seem to recall that the interface to IDE drives was fairly stateful. All of that needs to maintained in a separate dump driver.

Plus there is the challenge that many high-end systems where people want to employ this kind of thing have SCSI disks. At least that has been my impression.

Using kexec to switch to an independent program to do the dump sounds sane. If you have a simple system with just IDE you could run a very minimal 4K IDE dumper. If you have a full sized system you can use a kernel+initrd and do all kinds of interesting things. Dump to disk. Dump to network etc. Stop and work interactively with the user to examine the state of the machine when it crashed.

Basically you are only limited by how much of your system your monitor can put into a sane state.

Plus except for a small stub the code with kernel+initrd is almost totally user space and using the normal path for kernel drivers. The kernel drivers that are used may need to be made a little more robust but that is not a big deal.

What is most attractive is using kexec to put the policy in user space seems to be the unix way.

Eric

Ok, That is fairly simple, and can be implemented easily, so it is a good place to start.

A buffer to save old kernel state is needed. At the very least we need to save the old register state, copying over several megabytes of data won't hurt the picture.

It has a little more exposure to on-going DMA transfers than running from a reserved area of memory, as some of that memory may have been setup as DMA buffers by the dying kernel.

If it turns out that on-going DMA is a problem I see that as something we can fix later on.

Yes I would like to look at that.

I am tempted to suggest the data buffer with the registers and memory actually be in ELF format with virtual address representing where the data was in memory, and the physical addresses reporting where the data actually is in memory now. Then we could just grab everything with /dev/mem..

But I don't know how much of pain this would be.

Eric

/dev/mem expects mem_map to be there, the size of which would easily blow away the reserved section. /dev/oldmem (or whatever we call it) is just a magic copy that does a kmap-like operation to get at pages without a struct page.

M.

Don't we already do that for the mmap case so we can access I/O devices?

But I agree if there is a dependence there having a /dev/rawmem thing that does not need a page struct would make sense. I just don't think we actually need it...

Eric

Right, and some platforms already support this, fortunately.

I think this could end up being a good thing. It gives more people a stake in making sure that driver shutdown() routines work well.

Jesse

Which actually is one of the items open for discussion currently. For kexec on panic do we want to run the shutdown() routines?

Eric

We'll have to do something about incoming dma traffic and other stuff that the devices might be doing. Maybe a arch specific callout to do some chipset stuff?

Jesse

I vote for sleeping for 5 seconds ;-) Should kill off most of it ...

M.

Wake up smell the coffee.

- Bus masters that run forever
- Devices that need to flush before reset is asserted (eg IDE disk)

...

Does ongoing DMA actually matter? After all,the memory which is being dma-ed into is pre-reserved and allocated for that purpose, and the dump kernel won't be using it.

It would be polite to pause for a number of seconds to allow things to go quiet, but apart from that I think all we need to ensure is that the drivers in the dump kernel firmly whack the hardware before reinitialising it?

If we can ensure the addresses where the new kernel will run will never have DMA pointed at them I actually don't think so. This is why last year I recommended building a kernel that runs at a non-default address and finding a way to simply preload it there.

Eric

We DMA into arbitary allocated pages anywhere in the memory space, so you never know where is safe other than areas preallocated during the old kernel run.

Since you can just clear the master bit on each PCI device it isnt a big deal to protect against. (except a couple of devices that forget to honour it)

Alan I just reread what you said and it appears we are in violent agreement about the facts.

Different methods but...

Or those devices that hang the machine when you clear it. Or the ioapics which loose the ability to generate interrupts when you clear the master bit, and with the i82559 timer behind them you can't get your new kernel to boot.

Plus there are all of the non-pci devices.

And there is the fact that the pci configuration access methods are frequently BIOS calls.

So I do see just clearing the master bit on each PCI devices to as dangerous as calling the shutdown methods.

Eric

There are none. Its required by the PCI spec and used by BIOS vendors during the boot sequence. So its a *tested* approach.

You will be running bios code on some systems every time you read the cmos clock, every time you touch pci config space, every time you hit a key, even in your new kernel boot up path - whats your point

Then we violently disagree

Enabling is required. Clearing is not. The particular instance I was thinking of was disabling memory access and leaving I/O enabled.

Only that in many instances BIOS code can do things we don't expect. And when we start out with the machine in an unknown state the risk is worse.

yes.

Eric

Probably or you may not be able to recover some devices. In some cases simply turning off the master bit might be enough however.

We discussed this at kernel summit a bit - it'd be safer to make the devices clear down on boot up, rather than shutdown, if possible ... less work to do on the unstable base.

Maybe we could shut down the devices on bringup, then bring it up again (no I'm not kidding ;-)) ... should reuse the code.

M.

Agreed, but I think for starters we should capture the low hanging fruit by calling the shutdown method and then increasingly harden the code by performing less in the kernel that panics and more in the cleanup kernel.

That way we can concentrate first on the interfaces to the rest of the kernel. And then we can make the solution bullet proof.

Might not be a bad idea.

Eric

We really don't want to be calling driver shutdown functions from a crashed kernel.

Then at the very least you need to disable bus mastering and have specialist recovery functions for problematic devices. The bus mastering one is essential otherwise bus masters will continue to DMA random data into your new universe.

Other stuff like graphics cards and IDE may need care too.

But they're welcome to do that: the memory for the DMA transfer has already been allocated and our new universe will not be touching it.

What we need to do is to ensure that the new kexec-ed kernel appropriately whacks the devices to stop any in-progress operations. So it's the probe() and open() routines which need to get the device into a sane state, not the shutdown routines.

This way:

- We have less devices to take care of: we only care about those devices which are needed for a successful dump.

- We are poking at these devices in a known-good kernel, not from within a kernel which has wrecked itself.

- Any devices which are performing DMA to/from the old kernel's memory can just keep on doing that. The new kernel doesn't care, unless it needs those devices for dumping.

Yeah, for the most part that should be ok. We can be paranoid about misdirected DMA later... (Some platforms will let you protect memory regions at the chipset level, so it seems like the new kernel should be so protected until it's actually jumped to by kexec, but prior to unprotecting it you'd want to make sure that a bad DMA from the broken kernel doesn't hose it.)

That makes sense, and means that drivers may want to call a shutdown-like routine in their probe functions to make sure their device is in a known state before starting. But all of this is very driver specific it seems.

Jesse

Alan if we call anything the shutdown methods really are the thing to call. Because they are exactly the specialty recovery functions for problematic devices.

Of course no matter what we do will this work 100% of the time because part of what we will be fighting is broken hardware. However we should be able to get a mechanism that works most of the time.

What is your concern with stopping DMA?
- Not smashing the recovery routine.
- Getting a corrupted core dump because of on-going DMA?

Eric

Shutdown methods will typically call into the slab allocator and the page allocator to free stuff, and they are pretty common sources of oopses. Often with locks held. You run an excellent change of deadlocking.

Possibly one could add

#ifdef CONFIG_WHATEVER
	if (unlikely(oops_in_progress))
		return;

#endif

to the relevant entry points.

The shutdown routines may also call into sysfs/kobject/procfs release entry points, and they're even more popular oops sites.

We really want to get into the new kernel ASAP and clean stuff up from in there.

Hmm.. Last I looked shutdown methods typically don't exist at all. The shutdown methods are explicitly separated from the remove methods for exactly this reason. It is a BUG for any shutdown method to free memory. Their only function is to shutdown the hardware.

Again. If a shutdown method does that it is a BUG. Only the remove method should do that.

If I actually believed that shutdown methods existed, and did that I would be in favor of writing a patch to test for any accesses of memory management or sysfs/kobject/procfs release stuff and BUG if it happened.

I agree. However the gymnastics for doing that have not been worked out. The drivers cannot clean up stuff yet, nor do we have a good way to run in memory where DMA transfers on not ongoing.

So for a first pass I think calling the shutdown methods make sense.

For a second pass we need to use a relocatable that can do everything itself. And we should run it out of a reserved area of memory.

But the first pass is worth it (at least in the kexec tree) to sort out all of the interface issues and catch the low hanging fruit.

Eric

OK. But some (most) of them will sleep, too. And we shouldn't sleep in a dead kernel.

Don't we? The 16M of memory was allocated up-front at kexec load time[*], so nobody will be pointing DMA hardware at it. And the dump kernel won't be pointing DMA hardware at the crashed kernel's pages.

Well. There aren't any.

A significant proportion of kernel crashes happen from [soft]irq context, from which we cannot call shutdown methods. So we need to be able to bring up the dump kernel without having run driver shutdown functions anwyay..

[*] At least, I _assume_ the 16MB will be prereserved,
physically-contiguous and wholly within ZONE_NORMAL. Is this wrong?

Probably not. And that is legitimate...

No but we will be running in the first 16M of memory. The 16M that is allocated is currently used to hold a copy of the low 16M.

Which makes them both safe and worthless. On the normal kexec path they we will need to get them written though.

Well if calling shutdown is not really usable, then I we had better transition quickly beyond using it...

The problem is that we really won't be using it for running code out of because of i386 kernel limitations. Unless someone can tell my why 0 -16MB won't have DMA traffic in them. Or how to run a kernel at an address other than 1MB.

I suspect we can play with the initial page tables and how virtual addresses map to physical addresses and fairly simply generate a relocatable kernel. I have not had a chance to investigate that though. Once we have that it will be trivial to run out of the reserved 16M and many of the practical problems melt away.

Eric

IIRC, what Adam did is to relocate the bottom 16MB of mem into the reserved buffer and execute into the bottom 16MB. Yes, that probably does leave some DMA issues that we should fix up as you suggest above, but I think it's good enough for a first pass at the problem.

M.

I guess I don't have a problem with that as long as I don't have to chase the bugs.

If I do have to chase the bugs I would rather call the shutdown methods and just say things don't _yet_ work in cases where it is not safe to call them.

I guess I just like to be easily to explain what does not work _yet_.

With running at the same addresses we also have a rogue cpu problem as well. If we don't kill the cpu before the new kernel starts what happens if it starts executing some random bit of the new code...

Eric

Probably. I have witnessed network RX causing memory corruption, before the kexec code started downing the network interfaces on the user space side. I suspect data capture from sound cards or video capture cards would have the same issue.

The way I have observed this in the past is to kexec memtest86, on a machine with known good memory, and then attempt to ping it :)

What especially worries me about the low 16MB is that it is the DMA zone for ISA devices. Old sound cards in particular. Most of that is output but....

Eric

As long as the "init" routines are run on every startup (not just kexec ones), they should get plenty of testing (though not from bad card state).

I still think we could share code by running the shutdown routines from the *new* kernel before trying to init the card if they're written in a robust way so as to allow it ... is that insane?

M.

And I know for a fact that many init routines won't initialize a card in a bad state currently. That is my most frequent failure in the normal kexec case, when things are not in a

As a rough feel yes that is sane. Redundant but sane. I would like to hear what greg thinks of it though.

Eric

Oh, yes, I know that ... I'm just saying we should fix it ;-)

M.

Besides: when I want a crash dump, I want to dump the system state as of the moment it crashed and not as of $RANDOM_STUFF later, well-intentioned or not.

Let's reset the devices to a known state from within that kernel. If it can't do that for some reason, it should reboot the system through the BIOS (which had better be able to do it), and we're no worse off than before.

Matthias Urlichs

Completely random happenings occurring when they are trivial to avoid. Given all the worries about SHA signed in kernel standalone objects I find it farcical that the same people don't even care about ensuring something isnt DMAing over their dump partition description.

Alan

eh? People do care. The point here is that we should stop the DMA in the dump kernel, not from within the broken kernel.

btw, if we simply insert a five-second-pause, what problems does that leave? Network Rx, which is OK. Disk writes will have completed (?). What remains?

And pray just how do you expect to prove that the dump kernel isnt being overwritten *as* it is being loaded.

Network RX is the obvious one since we've no idea where the DMA is going in memory.

It was preloaded.

Of course, there's an assumption here that the dead kernel doesn't scribble on pages which were never available to its page allocator. If DMA somehow goes off and scribbles on the dump kernel we lose.

See above. We assume that network RX DMA won't be scribbling in the 16MB which was pre-reserved. That's reasonable. We _have_ to assume that.

If the new kernel image starts with an SHA hash check including the SHA hash check code that can be pretty robust as a sanity check.

Ok

Okay, I may be confused a bit but I *thought* kexec was going to load the thin, new kernel (e.g. read from disk operations, which is better than write to disk operations from the sick kernel).

This concept of having it pre-loaded sounds interesting, protecting it from being written on doesn't bother me much, but why *not* read it from disk/filesystem and then use the SHA hash in the newly loaded & exec'd kernel to make sure that what we loaded was sane?

That sounds simpler than changing the kernel load process around, ensuring you have the new kexec'd kernel build and loaded, etc. At least it sounds simpler and more in line with using kexec for fastboot as well.

gerrit

/sbin/kexec will load it with sys_kexec_load, before the kernel becomes sick.

Exactly. That is where the SHA hash and all of the features will go in the new ``kernel''. What we are exec is an arbitrary stand-alone program. I suspect a SHA hash generator and checker is something we can easily add as a wrapper.

The only process that is going to be changed around is where we store the kernel before we transfer control to it, and when/and how that transfer of control happens.

The beauty of kexec is all of these fun things become user problems from the point of the view of the sick kernel so it does not need to worry about them.

I will be happy to see a SHA patch for /sbin/kexec.

Eric

crypto/sha1.c provides all the code you need.

Yep that is the easy part, finding a sha1 implementation. The interesting part is the logic that hooks in the code and computes and checks the hash. Especially with the area the sha1 code checkes including the sha1 check code :)

Eric

...

This is what I had missed - the seperation of load and exec.

Thanks, Eric!

gerrit

If you wait a few seconds before verifying that the checksum of the rescue kernel is still correct, then you should be able to be reasonably sure that there won't be any corruption.

Nothing's 100% here, of course. But the chances that a delayed network DMA causes the rescue kernel to write its dump data to an area it shouldn't write to are small enough not to matter. IMHO.

Matthias Urlichs

Also audio and video capture on cyclic buffers can theorically go on forever sending irqs from time to time while they're at it.

OG.

I was asking so I could give a better answer.

As far as I can tell the long term solution is to simply run the dumper from memory that reserved in the kernel that called panic().

At that point you might get a corrupt DUMP. But it is extremely unlikely any DMA transactions will touch your reserved memory. Only the most buggy drivers or hardware will be running DMA to the invalid addresses. At which point there is very little you can do in any event.

Eric

Nothing we do is ever going to give us 100% perfect crashdumps from any given situation. I think we need to just accept that and take care not to destroy anything important (ie disk data), and try to get the best info into the dump we can.

M