as_area_destroy() broken on ppc32

[Martin Decky]

The implementation of as_area_destroy() seems to be broken on ppc32. The framebuffer server is killed due to various memory management problems. The user space heap allocator has been extensively examined and tested on other platforms and it doesn't seem to be the culprit.

The problem can be worked around by removing the memory sharing code in the framebuffer server. The problem can be completely worked around by changing the kernel function sys_as_area_destroy() into a no-op.

[Martin Decky]

Fixed in mainline,987 for ppc32. Initially, the Page Hash Table was used more like a huge TLB on ppc32. However, this is slightly misguided since the PHT is closer to Translation Storage Buffer on sparc64 (except that it is mandatory, while TLB is optional on ppc32). Currently the PHT gets invalidated correctly in as_area_destroy() and friends (although the current solution is suboptimal).

[Jakub Jermář]

Replying to decky:

Fixed in mainline,987 for ppc32. Initially, the Page Hash Table was used more like a huge TLB on ppc32. However, this is slightly misguided since the PHT is closer to Translation Storage Buffer on sparc64 (except that it is mandatory, while TLB is optional on ppc32). Currently the PHT gets invalidated correctly in as_area_destroy() and friends (although the current solution is suboptimal).

So what was the actual problem? IMO, it should not matter much whether the PHT is viewed either as an TLB or as an TSB. Note that the two are being invalidated together inside the TLB shootdown critical section (or very close to each other).

[Martin Decky]

So what was the actual problem? IMO, it should not matter much whether the PHT is viewed either as an TLB or as an TSB. Note that the two are being invalidated together inside the TLB shootdown critical section (or very close to each other).

The problem was that the PHT was not invalidated at all in this particular case, only the real TLB was.

[Jakub Jermář]

The problem was that the PHT was not invalidated at all in this particular case, only the real TLB was.

Oh, ok. I see. tlb_invalidate_pages() and tlb_invalidate_all() didn't do the right thing, whereas tlb_invalidate_asid() did.

[Martin Decky]

The ticket has been edited to talk only about the ppc32 issue, which has been already solved. A separate ticket for mips32 will be created.