busdma_iommu: map without extra offset bytes

[dougm]

Summary

iommu_map takes offset and size parameters, and promises to allocate a range that satisfies boundary constraints. iommu_bus_dmamap_load_something1 adds offset to size before calling iommu_map, and shrinks the result by offset, perhaps to work around a bug that was in iommu_map before commit 11fced21ccea1b80327d159a4c27046cb1f46952. This change removes that addition and subtraction of offset to avoid requiring a larger address range than necessary.

Test Plan

Peter, a run with DMAR turned on, please.

[kib]

To test, ask Peter Holm to do a run with DMAR turned on.

sys/dev/iommu/busdma_iommu.c

619

I agree with the removal of the asserts and the clip above this one, but IMO tag maxsegsz check should be kept, at least in the form of the assert.

[dougm]

sys/dev/iommu/busdma_iommu.c

619

The maxsegsz assert is still there, at line 639(before)/605(after).

[pho]

https://people.freebsd.org/~pho/stress/log/log0314.txt

[dougm]

iommu_map->iommu_gas_map->iommu_gas_find_space requires that the size parameter be a multiple of page size, so restore the round-up to respect that.

[pho]

panic: segment too large: ctx 0xfffff800041ca180 start 0xfef01000 end 0xfef24000 buflen1 0x23000 maxsegsz 0x22400

https://people.freebsd.org/~pho/stress/log/log0316.txt

[dougm]

I guess rounding buflen1 up to a multiple of page size, when maxsegsz is not a multiple of page size, is a problem. So restore check to reduce buflen1 to maxsegsz.

[pho]

Kernel page fault with the following non-sleepable locks held:
exclusive rw vm object (vm object) r = 0 (0xfffff80004837b58) locked @ x86/iommu/intel_idpgtbl.c:550
exclusive sleep mutex AHCI channel lock (AHCI channel lock) r = 0 (0xfffffe003ce28400) locked @ kern/kern_mutex.c:211
stack backtrace:
#0 0xffffffff80c85445 at witness_debugger+0x65
#1 0xffffffff80c8659a at witness_warn+0x3ea
#2 0xffffffff810fcce6 at trap_pfault+0x86
#3 0xffffffff810cdc18 at calltrap+0x8
#4 0xffffffff81078e7e at iommu_gas_map+0x15e
#5 0xffffffff81077339 at iommu_bus_dmamap_load_something+0x119
#6 0xffffffff81076995 at iommu_bus_dmamap_load_buffer+0x1c5
#7 0xffffffff80c55a3e at _bus_dmamap_load_ccb+0x20e
#8 0xffffffff80c557cc at bus_dmamap_load_ccb+0x8c
#9 0xffffffff803929d9 at xpt_run_devq+0x2f9
#10 0xffffffff80395de7 at xpt_release_simq+0x67
#11 0xffffffff80c3027a at softclock_call_cc+0x15a
#12 0xffffffff80c31b96 at softclock_thread+0xc6
#13 0xffffffff80bc9850 at fork_exit+0x80
#14 0xffffffff810cec8e at fork_trampoline+0xe

https://people.freebsd.org/~pho/stress/log/log0317.txt

[dougm]

Remove 'offset' in computing initial size, but not when when examining iommu_map product.

[dougm]

Recover missing "- offset".

[dougm]

Remove the material part of this patch, because a stealth reviewer finds a bug in it, and leave only the style bits which do simplify some obscure code.

[alc]

sys/dev/iommu/busdma_iommu.c

602–603

I would recommend adding a comment here to say that this is handling a split buffer.

[dougm]

Add recommended comment.