Add CO-RE support for kernel modules #1300

brycekahle · 2024-01-08T20:41:05Z

Adds KernelModule() (string, error) to ProgramSpec. This function attempts to determine the kernel module for kprobe/fentry programs, by parsing /proc/kallsyms.
Adds KernelModuleTypes map[string]*btf.Spec to ProgramOptions, so users can provide their own kmod BTF.

The usage of a *Spec in parts of CO-RE had to be abstracted to avoid copying/merging of []Type slices.

This does work (tested with nf_conntrack kmod and __nf_conntrack_hash_insert kprobe), but I imagine you have thoughts about how to adapt this approach.

brycekahle · 2024-01-08T20:51:48Z

It looks like some examples are failing because tracefs is not mounted. I could make KernelModule always return "" in face of an error, but I'm not sure that is a good idea.

lmb

Nice, I was worried that this would be a lot more complicated!

I still have my old questions re semantics and corner cases:

As mentioned, do we have to deal with module reload / unload? It would invalidate the list of probeable functions for sure. It might change kmod BTF as well.
Using the fn name to figure out the kmod seems a bit error prone.
You propose that CO-RE for a probe attached to a kprobe in a kernel module will look at set of vmlinux types + set of containing kmod types. What about a kprobe on a vmlinux function which needs access to kmod types? I'm thinking of data behind void *ctx and similar. (There is also a related case of probe on kmod which needs access to other kmod types, but that is probably rare.)

For the implementation, I was thinking that instead of introducing mergedSpec we'd operate on []*Spec. ProgramOptions.KernelTypes would become []*Spec as well. CORERelocate would iterate the []*Spec and try to find a valid set of relocations for each target. The logic to find the best possible match would have to be adjusted accordingly:

ebpf/btf/core.go

Lines 276 to 294 in a8be855

    
           if score > bestScore { 
        
           	// We have a better target already, ignore this one. 
        
           	continue 
        
           } 
        
           if score < bestScore { 
        
           	// This is the best target yet, use it. 
        
           	bestScore = score 
        
           	bestFixups = fixups 
        
           	continue 
        
           } 
        
           // Some other target has the same score as the current one. Make sure 
        
           // the fixups agree with each other. 
        
           for i, fixup := range bestFixups { 
        
           	if !fixup.equal(fixups[i]) { 
        
           		return nil, fmt.Errorf("%s: multiple types match: %w", fixup.kind, errAmbiguousRelocation) 
        
           	} 
        
           }

This also decouples the CO-RE logic from the decision which set of types to use.

elf_reader.go

internal/tracefs/kprobe.go

prog.go

btf/btf.go

btf/core.go

brycekahle · 2024-01-12T17:42:30Z

For the implementation, I was thinking that instead of introducing mergedSpec we'd operate on []*Spec. ProgramOptions.KernelTypes would become []*Spec as well. CORERelocate would iterate the []*Spec and try to find a valid set of relocations for each target.

Since the type IDs of kernel modules conflict with each other, you are only ever going to be dealing with vmlinux plus one optional kmod BTF. libbpf makes this same assumption.

brycekahle · 2024-01-12T17:45:21Z

You propose that CO-RE for a probe attached to a kprobe in a kernel module will look at set of vmlinux types + set of containing kmod types. What about a kprobe on a vmlinux function which needs access to kmod types?

With the changes you requested, you could accomplish this by setting KernelModule in the program spec (and supplying KernelModuleTypes in program options, if you want). That field doesn't affect the attachment in any way, just the BTF load and CO-RE relocation.

brycekahle · 2024-01-12T17:52:34Z

do we have to deal with module reload / unload?

I don't see a way to know if a module has been unloaded and then loaded with different contents, other than maybe comparing sizes from /proc/modules. This would be unlikely for a kernel module in the source tree right?

lmb

Okay, went over the PR again. Here is what I would propose, let me know what you think.

Skip ProgramSpec.KernelModule. Instead, ProgramOptions.KernelTypes becomes []*btf.Spec as discussed. If KernelTypes is nil, the library tries to find the appropriate BTF to use, via means of parsing kallsyms / available_filter_functions and doing a lookup of ProgramSpec.AttachTo. Most of the time this will be just vmlinux. Some of the time this will be vmlinux + kmod via LoadKernelModuleSpec.
kallsyms / available_filter_functions: which one is more appropriate to use? I'd also prefer to not export KernelModule for now, since it ties us into an API. Probably best to stick this into internal/kallsyms. Need to figure out whether this needs to be cached, and if yes how to invalidate the cache when the set of loaded modules changes. Something in proc that lists loaded modules maybe?
Merge FlushKernelModuleCache into btf.FlushKernelSpec(). There should be a single function to clean up cached info, and in a way this is all related to btf stuff anyways. We could rename the function if need be.
CORERelocate will work against []*Spec as discussed. I think it makes sense to reject target type ID lookups for kmods as long as there is no canonical way to refer to these IDs from the kernel side. This might need some finessing in CORERelocate, I can probably take a look at that once the first implementation is in.

prog_test.go

lmb

Okay, went over the PR again. Here is what I would propose, let me know what you think.

Skip ProgramSpec.KernelModule. Instead, ProgramOptions.KernelTypes becomes []*btf.Spec as discussed. If KernelTypes is nil, the library tries to find the appropriate BTF to use, via means of parsing kallsyms / available_filter_functions and doing a lookup of ProgramSpec.AttachTo. Most of the time this will be just vmlinux. Some of the time this will be vmlinux + kmod via LoadKernelModuleSpec.
kallsyms / available_filter_functions: which one is more appropriate to use? I'd also prefer to not export KernelModule for now, since it ties us into an API. Probably best to stick this into internal/kallsyms. Need to figure out whether this needs to be cached, and if yes how to invalidate the cache when the set of loaded modules changes. Something in proc that lists loaded modules maybe?
Merge FlushKernelModuleCache into btf.FlushKernelSpec(). There should be a single function to clean up cached info, and in a way this is all related to btf stuff anyways. We could rename the function if need be.
CORERelocate will work against []*Spec as discussed. I think it makes sense to reject target type ID lookups for kmods as long as there is no canonical way to refer to these IDs from the kernel side. This might need some finessing in CORERelocate, I can probably take a look at that once the first implementation is in.

brycekahle · 2024-01-16T18:51:06Z

kallsyms / available_filter_functions: which one is more appropriate to use

It was hard to find a conclusive comparison between the two based on the kernel source, but kallsyms is definitely more inclusive than available_filter_functions. available_filter_functions is from the ftrace subsystem.

brycekahle · 2024-01-16T18:54:23Z

I'd also prefer to not export KernelModule for now, since it ties us into an API

In the case of btfhub-supported kernels, a user will need a way to know if/which module a function lies within. That way we can lookup the appropriate BTF to pass into the ProgramOptions.

Since the library would already have this functionality, why not export it? Otherwise every user of btfhub, kmods, and cilium/ebpf will need to duplicate it.

brycekahle · 2024-01-17T17:57:01Z

Merge FlushKernelModuleCache into btf.FlushKernelSpec(). There should be a single function to clean up cached info, and in a way this is all related to btf stuff anyways

FlushKernelModuleCache is flushing the cache of kernel function -> module name, not BTF related.

brycekahle · 2024-01-17T18:40:02Z

If KernelTypes is nil, the library tries to find the appropriate BTF to use, via means of parsing kallsyms / available_filter_functions and doing a lookup of ProgramSpec.AttachTo. Most of the time this will be just vmlinux. Some of the time this will be vmlinux + kmod via LoadKernelModuleSpec.

This moves the loading of kernel/kmod BTF up from CORERelocate (which is a public func) to newProgramWithOptions because it needs access to ProgramSpec.AttachTo. That seems like a breaking change if folks expect CORERelocate to load kernel BTF for them.

brycekahle · 2024-01-17T19:17:05Z

I made some changes that are close to what you want, but had to fill in some gaps. I would like to see KernelModule function exported.

brycekahle · 2024-01-22T21:40:16Z

@lmb can you take another look?

brycekahle · 2024-01-23T21:01:44Z

Ugh. Since we need different KernelTypes []*btf.Spec per-program, it kind of excludes us from using NewCollection/NewCollectionWithOptions because that only takes one set of options for all the programs.

One idea would be KernelTypes map[string]*btf.Spec with either a sentinel or empty string value for the base kernel BTF. This would allow you to pass all the possibly needed BTF for a collection. Downsides:

both the user and library would have to do the func->kmod lookup for each program.
the map only makes sense for a collection, not for a single program, so it would be in CollectionOptions

Another idea is a callback function func(kmodName string) (*btf.Spec, error) that the library could call if specified AND it was unable to find default BTF.

eiffel-fl

Hi!

Thank you for this contribution!
We are interested by using it in Inspektor Gadget.
So far, I only have one comment but I will taker another look later to get a broader understanding.

Best regards.

btf/btf.go

lmb

Going back to semantics: your current implementation means that a type defined in a kernel module can "shadow" a vmlinux type if it's better according to our scoring. I'm on the fence whether it'd be better to always prefer vmlinux if there is a matching target, no matter how good a kmod type is. Do you have an opinion on that?

internal/kallsyms/kallsyms_test.go

btf/btf.go

btf/core.go

btf/btf.go

lmb · 2024-01-24T12:54:13Z

btf/core.go

 	for localType, group := range relosByType {
 		localTypeName := localType.TypeName()
 		if localTypeName == "" {
 			return nil, fmt.Errorf("relocate unnamed or anonymous type %s: %w", localType, ErrNotSupported)
 		}

-		targets := target.namedTypes[newEssentialName(localTypeName)]
-		fixups, err := coreCalculateFixups(group.relos, target, targets, bo)
+		fixups, err := coreCalculateFixups(group.relos, &mergeTarget, mergeTarget.NamedTypesIterate(newEssentialName(localTypeName)), bo)


Merging specs up front is going to be more complicated with lazy copy going in. How about the following pseudo go:

for localType, group := range relosByType { var bestFixup for _, target := range targets { fixup, err := coreCalculateFixup(...) if fixup > bestFixup { bestFixup = fixup // other checks } } if bestFixup == nil { // generate poison all } }

Instead of merging the specs, we add a loop here which iterates over targets and tries to find a valid fixup.

We factor out the code to do scoring and call if from here in addition to coreCalculateFixups

ebpf/btf/core.go

Lines 276 to 294 in f95957d

if score > bestScore {

// We have a better target already, ignore this one.

continue

}

if score < bestScore {

// This is the best target yet, use it.

bestScore = score

bestFixups = fixups

continue

}

// Some other target has the same score as the current one. Make sure

// the fixups agree with each other.

for i, fixup := range bestFixups {

if !fixup.equal(fixups[i]) {

return nil, fmt.Errorf("%s: multiple types match: %w", fixup.kind, errAmbiguousRelocation)

}

}

We move generating the poison all fixup to after the loop over targets in this function:

ebpf/btf/core.go

Lines 298 to 309 in f95957d

// Nothing at all matched, probably because there are no suitable

// targets at all.

//

// Poison everything except checksForExistence.

bestFixups = make([]COREFixup, len(relos))

for i, relo := range relos {

if relo.kind.checksForExistence() {

bestFixups[i] = COREFixup{kind: relo.kind, local: 1, target: 0}

} else {

bestFixups[i] = COREFixup{kind: relo.kind, poison: true}

}

}

I'm not actually merging them, just storing pointers to each Spec and iterating through them with wrapper functions. I updated to handle your changes in main. Let me know what you think.

Okay, works for me. I'm not a big fan of mergedSpec, it seems too special purpose for me. Can you move it to core.go so that it's close to where it'll be used?

lmb · 2024-01-24T13:10:41Z

Ugh. Since we need different KernelTypes []*btf.Spec per-program, it kind of excludes us from using NewCollection/NewCollectionWithOptions because that only takes one set of options for all the programs.

Ah, that is annoying :( Why do you want to avoid multiple NewCollection calls?

One idea would be KernelTypes map[string]*btf.Spec with either a sentinel or empty string value for the base kernel BTF.

That could work. I'd probably use vmlinux as a sentinel. OTOH we could also keep the current KernelTypes and add KernelModuleTypes map[string]*Spec, which means we don't break that part of the API.

both the user and library would have to do the func->kmod lookup for each program.

This is because you don't want to load all kmod by default? What if you only parsed kmod for all loaded modules instead?

the map only makes sense for a collection, not for a single program, so it would be in CollectionOptions

Hmm, I don't follow, sorry. We could still stick the map in ProgramOptions I think? Not sure how we would make use of it when loading ebpf.Map for example.

Another idea is a callback function func(kmodName string) (*btf.Spec, error)

I'm a bit wary of callbacks. If we add this it begs the question whether we'd use this for vmlinux as well? Does it deprecate KernelTypes? Gut feeling only, but I'd prefer to not use a callback.

brycekahle · 2024-01-24T16:49:12Z

Why do you want to avoid multiple NewCollection calls?

We have a bunch of eBPF programs in the same object file. That is currently a single collection. Anything that needs kernel modules would need to be extracted out and dealt with separately, probably as individual program loads. That makes shared map handling way more complicated. It isn't impossible, but is a smell about the API maybe being wrong.

brycekahle · 2024-01-24T16:53:25Z

This is because you don't want to load all kmod by default? What if you only parsed kmod for all loaded modules instead?

If you mean modules already loaded in the kernel, that can still be 100 or more. That is a lot of CPU and memory used for probably no reason.

Otherwise, we have to loop through the kprobe functions, determine which modules are needed, load those BTF. The library when loading each kprobe would also need to do that lookup of function to kmod, to load the correct BTF from the map.

brycekahle · 2024-01-24T16:54:19Z

Hmm, I don't follow, sorry. We could still stick the map in ProgramOptions I think?

We could put it in ProgramOptions but it doesn't make a lot of sense, because a single eBPF program can have max 2 BTFs, the kernel and a kernel module.

brycekahle · 2024-01-24T16:55:08Z

your current implementation means that a type defined in a kernel module can "shadow" a vmlinux type if it's better according to our scoring.

How would that happen? kmod BTF is generated using the vmlinux BTF so that it doesn't include anything already in vmlinux BTF.

brycekahle · 2024-01-30T19:52:42Z

@lmb updated using a map for the module types.

lmb · 2024-02-02T13:38:56Z

Sorry for the long radio silence. I've been busy hashing out cilium/cilium work.

@brycekahle

We could put it in ProgramOptions but it doesn't make a lot of sense, because a single eBPF program can have max 2 BTFs, the kernel and a kernel module.

I don't see the problem to be honest. ProgramOptions is already not scoped to a specific program.

How would that happen? kmod BTF is generated using the vmlinux BTF so that it doesn't include anything already in vmlinux BTF.

By using distinct types with the same name in multiple compilation units. Try grepping for ring_buffer in the kernel sources.

Which ring_buffer do we prefer? The kmod one? Or the kernel? Which one of the kernel ones?

It is worth noting that I don't think libbpf can accomplish [relocating types from multiple kmod] at the moment.

Does that mean we can't do better? ;P

I don't think we could automagically know which additional modules to load BTF for, so the user would have to provide/specify that somehow.

We can't figure out which modules we need up front, but we could do something like the following:

Define an ordering in which we visit kmod (one reason I want []*Spec as argument to CORERelocate: it gives us an order)
Keep iterating specs until we've found a target for all relocations (logic for this will be tricky)

These are the correct semantics to me (rather than just vmlinux + kmod) but I understand that the immediate need is simpler. I still want to keep the door open for this later on though.

With kernel module dependencies if A depends on B, it appears that A's BTF contains all the necessary types it is using from B.

That's interesting! Can you show me an example or tell me how to reproduce that?

@alban

I'd prefer if [merged kmod btf] worked, but I understand we have to balance complexity and usefulness from real use cases.

Thanks for giving your view point and giving me confirmation that I'm not completely off xD

lmb

Looks fine, please make the changes to CORERelocate as we discussed. If targets is nil it can just return an error, which is what it was doing before.

I'm a bit concerned about the dead lock potential between kernel and kmod loading, I'd love there to only be a single lock. Oh well.

Finally, the bit I'm still missing is how you're going to use this with external kmod BTF?

You don't want to parse all loaded modules, since that might be in the 100s and could still be too slow.
There is no feedback mechanism for you to figure out which modules the lib wants.

I see two solutions: replace the map with the callback you proposed or expose programKernelModule on ProgramSpec. The latter seems fine, and would also allow dropping the export of ebpf.KernelModule. WDYT?

internal/kallsyms/kallsyms_test.go

lmb · 2024-02-02T14:09:53Z

btf/core.go

 	for localType, group := range relosByType {
 		localTypeName := localType.TypeName()
 		if localTypeName == "" {
 			return nil, fmt.Errorf("relocate unnamed or anonymous type %s: %w", localType, ErrNotSupported)
 		}

-		targets := target.namedTypes[newEssentialName(localTypeName)]
-		fixups, err := coreCalculateFixups(group.relos, target, targets, bo)
+		fixups, err := coreCalculateFixups(group.relos, &mergeTarget, mergeTarget.NamedTypesIterate(newEssentialName(localTypeName)), bo)


Okay, works for me. I'm not a big fan of mergedSpec, it seems too special purpose for me. Can you move it to core.go so that it's close to where it'll be used?

brycekahle · 2024-02-07T21:19:29Z

Finally, the bit I'm still missing is how you're going to use this with external kmod BTF?

You don't want to parse all loaded modules, since that might be in the 100s and could still be too slow.

There is no feedback mechanism for you to figure out which modules the lib wants.

You can see how I've implemented it at the moment: DataDog/ebpf-manager@8b77c9a

brycekahle · 2024-02-07T21:21:16Z

If targets is nil it can just return an error, which is what it was doing before.

This is not what it was doing. It tries to auto-load the kernel spec:

ebpf/btf/core.go

Lines 173 to 179 in b3432fb

    
           if target == nil { 
        
           	var err error 
        
           	target, _, err = kernelSpec() 
        
           	if err != nil { 
        
           		return nil, fmt.Errorf("load kernel spec: %w", err) 
        
           	} 
        
           }

lmb · 2024-02-16T14:22:34Z

This is not what it was doing. It tries to auto-load the kernel spec:

I meant this:

ebpf/btf/core.go

Lines 169 to 170 in 4267cbc

    
           // Explicitly check for nil here since the argument used to be optional. 
        
           return nil, fmt.Errorf("target must be provided")

I'm looking at merging this but realised that you merged main into your branch. Can you please undo that and rebase + squash on top of main?

brycekahle · 2024-02-16T23:33:52Z

Can you please undo that and rebase + squash on top of main?

Done. Let me know if there are any other changes you'd like me to make.

Signed-off-by: Bryce Kahle <bryce.kahle@datadoghq.com>

Signed-off-by: Lorenz Bauer <lmb@isovalent.com>

No changes to the source code. Signed-off-by: Lorenz Bauer <lmb@isovalent.com>

Signed-off-by: Lorenz Bauer <lmb@isovalent.com>

This allows dropping the fallback parameter and allows further refactoring. Signed-off-by: Lorenz Bauer <lmb@isovalent.com>

Kernel and module BTF are currently kept in two separate global variables, with separate locking. This creates a deadlock hazard. Refactor the code to use a single lock, and get rid of a lot of indirection which is not necessary anymore because Spec.Copy is cheap. Signed-off-by: Lorenz Bauer <lmb@isovalent.com>

The function is already invoked via btf.FlushKernelSpec(). Signed-off-by: Lorenz Bauer <lmb@isovalent.com>

Signed-off-by: Lorenz Bauer <lmb@isovalent.com>

lmb

@brycekahle I ended up making the changes myself.

Don't export FlushKallsysm separately.
Don't pass kmodName to CORERelocate.
Remove the deadlock risk from having two locks protecting vmlinux and modules

I feel like these came up during the review before, but I'm not sure what happened.

Sorry it took so long to get this over the finish line!

lmb · 2024-02-22T18:28:20Z

Going to merge this tomorrow morning UK time.

eiffel-fl

Hi!

Thank you for the initial work and the polish!
I do not have a broad knowledge about it but I have one question with regard to locking and found a nit.

Best regards.

eiffel-fl · 2024-02-23T06:08:17Z

btf/kernel.go

+		kernelBTF.Lock()
+		defer kernelBTF.Unlock()


I am not sure to understand the locking here.
Particularly, I would just hold the write lock while writing kernelBTF.kernel, but I surely miss something here. Can you please shed some light?

Yeah, the code is a bit hard to follow. Reason we do it this way is that dropping the read lock and acquiring the write lock are not atomic. Another goroutine might have preempted us and populated kernelBTF.kernel.

eiffel-fl · 2024-02-23T06:09:44Z

btf/kernel.go

+	base, err := LoadKernelSpec()
+	if err != nil {
+		return nil, fmt.Errorf("load kernel spec: %w", err)
+	}


This snippet can be moved right before you need it, i.e. line 86.

That would cause a deadlock: we already hold kernelBTF.Lock at that point, so acquiring it in LoadKernelSpec will block indefinitely.

brycekahle assigned lmb Jan 8, 2024

github-actions bot added the breaking-change Changes exported API label Jan 8, 2024

lmb requested changes Jan 12, 2024

View reviewed changes

brycekahle force-pushed the bryce.kahle/kmod-co-re branch 2 times, most recently from b02c8ed to d01b1d5 Compare January 12, 2024 20:11

lmb reviewed Jan 16, 2024

View reviewed changes

prog_test.go Outdated Show resolved Hide resolved

lmb reviewed Jan 16, 2024

View reviewed changes

brycekahle force-pushed the bryce.kahle/kmod-co-re branch 2 times, most recently from 494679f to a20efa4 Compare January 17, 2024 19:25

alban mentioned this pull request Jan 23, 2024

trace exec upper_layer: fix BTF type size relocation inspektor-gadget/inspektor-gadget#2400

Merged

brycekahle force-pushed the bryce.kahle/kmod-co-re branch from 59a51a1 to 5a8c94d Compare January 23, 2024 21:04

brycekahle mentioned this pull request Jan 23, 2024

btfhub-archive does not contain BTF for kernel modules DataDog/btfhub#7

Closed

3 tasks

eiffel-fl reviewed Jan 24, 2024

View reviewed changes

btf/btf.go Outdated Show resolved Hide resolved

lmb reviewed Jan 24, 2024

View reviewed changes

brycekahle force-pushed the bryce.kahle/kmod-co-re branch from e914652 to 913f586 Compare January 30, 2024 19:51

lmb requested changes Feb 2, 2024

View reviewed changes

brycekahle force-pushed the bryce.kahle/kmod-co-re branch from fd2e4dd to ce7b452 Compare February 16, 2024 23:32

brycekahle marked this pull request as ready for review February 16, 2024 23:33

brycekahle requested a review from a team as a code owner February 16, 2024 23:33

support for kmod CO-RE

7f00a40

Signed-off-by: Bryce Kahle <bryce.kahle@datadoghq.com>

lmb force-pushed the bryce.kahle/kmod-co-re branch from ce7b452 to 5fd9ae5 Compare February 22, 2024 15:09

lmb added 7 commits February 22, 2024 15:21

btf: make callers pass targets to CORERelocate

b32a1c5

Signed-off-by: Lorenz Bauer <lmb@isovalent.com>

btf: move kernel spec parsing to separate file

173d741

No changes to the source code. Signed-off-by: Lorenz Bauer <lmb@isovalent.com>

btf: add a test for LoadKernelModuleSpec

8796b9e

Signed-off-by: Lorenz Bauer <lmb@isovalent.com>

btf: don't call into kernelSpec directly

27b03ef

This allows dropping the fallback parameter and allows further refactoring. Signed-off-by: Lorenz Bauer <lmb@isovalent.com>

ebpf: remove FlushKernelModuleCache

95f27b8

The function is already invoked via btf.FlushKernelSpec(). Signed-off-by: Lorenz Bauer <lmb@isovalent.com>

kallsyms: extend test

215a22a

Signed-off-by: Lorenz Bauer <lmb@isovalent.com>

lmb force-pushed the bryce.kahle/kmod-co-re branch from 5ad459e to 215a22a Compare February 22, 2024 15:21

lmb self-requested a review February 22, 2024 18:25

lmb approved these changes Feb 22, 2024

View reviewed changes

eiffel-fl reviewed Feb 23, 2024

View reviewed changes

eiffel-fl mentioned this pull request Feb 23, 2024

Use BTF to check block_rq_insert arguments inspektor-gadget/inspektor-gadget#2531

Merged

lmb merged commit 4ad27ae into cilium:main Feb 23, 2024

alban mentioned this pull request Apr 1, 2024

pkg/uprobetracer: use kfilefields to identify inodes inspektor-gadget/inspektor-gadget#2669

Merged

wenhuwang mentioned this pull request Apr 19, 2024

retina-agent pod initialization failed to reconcile plugin dropreason: field NfConntrackConfirm: program nf_conntrack_confirm microsoft/retina#246

Closed

brycekahle deleted the bryce.kahle/kmod-co-re branch July 15, 2024 16:06

	if score > bestScore {
	// We have a better target already, ignore this one.
	continue
	}

	if score < bestScore {
	// This is the best target yet, use it.
	bestScore = score
	bestFixups = fixups
	continue
	}

	// Some other target has the same score as the current one. Make sure
	// the fixups agree with each other.
	for i, fixup := range bestFixups {
	if !fixup.equal(fixups[i]) {
	return nil, fmt.Errorf("%s: multiple types match: %w", fixup.kind, errAmbiguousRelocation)
	}
	}

	// Nothing at all matched, probably because there are no suitable
	// targets at all.
	//
	// Poison everything except checksForExistence.
	bestFixups = make([]COREFixup, len(relos))
	for i, relo := range relos {
	if relo.kind.checksForExistence() {
	bestFixups[i] = COREFixup{kind: relo.kind, local: 1, target: 0}
	} else {
	bestFixups[i] = COREFixup{kind: relo.kind, poison: true}
	}
	}

Add CO-RE support for kernel modules #1300

Add CO-RE support for kernel modules #1300

Conversation

brycekahle commented Jan 8, 2024 • edited Loading

brycekahle commented Jan 8, 2024

lmb left a comment

Choose a reason for hiding this comment

brycekahle commented Jan 12, 2024

brycekahle commented Jan 12, 2024

brycekahle commented Jan 12, 2024

lmb left a comment

Choose a reason for hiding this comment

lmb left a comment

Choose a reason for hiding this comment

brycekahle commented Jan 16, 2024

brycekahle commented Jan 16, 2024 • edited Loading

brycekahle commented Jan 17, 2024

brycekahle commented Jan 17, 2024

brycekahle commented Jan 17, 2024

brycekahle commented Jan 22, 2024

brycekahle commented Jan 23, 2024 • edited Loading

eiffel-fl left a comment

Choose a reason for hiding this comment

lmb left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

lmb commented Jan 24, 2024

brycekahle commented Jan 24, 2024

brycekahle commented Jan 24, 2024 • edited Loading

brycekahle commented Jan 24, 2024

brycekahle commented Jan 24, 2024 • edited Loading

brycekahle commented Jan 30, 2024

lmb commented Feb 2, 2024

lmb left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

brycekahle commented Feb 7, 2024

brycekahle commented Feb 7, 2024

lmb commented Feb 16, 2024

brycekahle commented Feb 16, 2024

lmb left a comment

Choose a reason for hiding this comment

lmb commented Feb 22, 2024

eiffel-fl left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

brycekahle commented Jan 8, 2024 •

edited

Loading

brycekahle commented Jan 16, 2024 •

edited

Loading

brycekahle commented Jan 23, 2024 •

edited

Loading

brycekahle commented Jan 24, 2024 •

edited

Loading

brycekahle commented Jan 24, 2024 •

edited

Loading