issue with patch "x86: no CPA on iounmap"
This is wrt to x86 git commit f56d005d30342a45d8af2b75ecccc82200f09600
"x86: no CPA on iounmap"
This can use performance issue. When a GART driver unmaps a RAM page, which was mapped as UC, this commit will still retain UC attribute on the kernel identity mapping. This can cause mysterious performance issue if this freed page gets used by kernel later.
For now we should change the attribute during iounmap and in future PAT infrastructure will have necessary hooks to avoid the aliasing issues.
thanks,
suresh
Quoted 11 lines · Suresh B
This is wrt to x86 git commit
f56d005d30342a45d8af2b75ecccc82200f09600
"x86: no CPA on iounmap"This can use performance issue. When a GART driver unmaps a RAM page, which was mapped as UC, this commit will still retain UC attribute on the kernel identity mapping. This can cause mysterious performance issue if this freed page gets used by kernel later.
For now we should change the attribute during iounmap and in future PAT infrastructure will have necessary hooks to avoid the aliasing issues.
this is a hard one; because the flipside is someone ioremapping the same page twice (which is not impossible if there's 2 1k bars)... and then one of them gets iounmapped, which would turn the page cachable again.
Quoted 4 lines · Suresh B
This is wrt to x86 git commit
f56d005d30342a45d8af2b75ecccc82200f09600
"x86: no CPA on iounmap"This can use performance issue. When a GART driver unmaps a RAM page,
thinking about this some more...
afaik the gart driver doesn't use ioremap....
(and it does caching control explicitly, and sets its pages back to cached)
Quoted 4 lines · Arjan van de Ven
Quoted 3 lines · Suresh B
This is wrt to x86 git commit
f56d005d30342a45d8af2b75ecccc82200f09600
"x86: no CPA on iounmap"
Quoted 8 lines
Quoted 1 line
This can use performance issue. When a GART driver unmaps a RAM page,
thinking about this some more...
afaik the gart driver doesn't use ioremap....
(and it does caching control explicitly, and sets its pages back to cached)
there are many GART drivers, and the method used depends on the GART driver. The following GART drivers still use ioremap in one way or another:
drivers/char/agp/amd-k7-agp.c
drivers/char/agp/ati-agp.c
drivers/char/agp/generic.c
drivers/char/agp/sworks-agp.c
drivers/char/drm/radeon_cp.c
the method use is in all cases the same: they use __get_free_page() to pick up a general RAM page, they do SetPageReserved() and then they use ioremap_nocache() to map it non-cached, and then they also program the GART to access those pages.
when the GART code deinits, it does an iounmap() on those pages, unmaps it from the GART hardware itself, does a ClearPageReserved() and does __free_page() to put the page into the general page pool again. So Suresh is right: these pages are currently marked UC at this point and we need to mark them cacheable.
we could do this automatically in iounmap() upon seeing a page_is_ram() that has PageReserved set. Or we could stick in a set_memory_wb() into the deinit [and ioremap_nocache()-failure] sequence.
Since we treat PageReserved pages specially in ioremap() already [we allow them, despite them being listed in the e820 map], i think the more robust solution is to recognize them in iounmap() as well - this way it cannot be forgotten accidentally. (and UC pages in the buddy are _hard_ to notice after the fact) There is no aliasing danger i believe: IO bars should never be marked as general RAM in the e820.
Ingo
Quoted 21 lines · Ingo Molnar
Quoted 10 lines · Arjan van de Ven
Quoted 4 lines · Suresh B
This is wrt to x86 git commit
f56d005d30342a45d8af2b75ecccc82200f09600
"x86: no CPA on iounmap"This can use performance issue. When a GART driver unmaps a RAM page,
thinking about this some more...
afaik the gart driver doesn't use ioremap....
(and it does caching control explicitly, and sets its pages back to cached)
there are many GART drivers, and the method used depends on the GART driver. The following GART drivers still use ioremap in one way or another:
drivers/char/agp/amd-k7-agp.c
drivers/char/agp/ati-agp.c
drivers/char/agp/generic.c
drivers/char/agp/sworks-agp.c
drivers/char/drm/radeon_cp.c
Quoted 8 lines
the method use is in all cases the same: they use
__get_free_page()to pick up a general RAM page, they doSetPageReserved()and then they useioremap_nocache()to map it non-cached, and then they also program the GART to access those pages.when the GART code deinits, it does an
iounmap()on those pages, unmaps it from the GART hardware itself, does aClearPageReserved()and does__free_page()to put the page into the general page pool again. So Suresh is right: these pages are currently marked UC at this point and we need to mark them cacheable.we could do this automatically in
iounmap()upon seeing apage_is_ram()that has PageReserved set. Or we could stick in aset_memory_wb()into the deinit [and ioremap_nocache()-failure] sequence.Since we treat PageReserved pages specially in
ioremap()already [we allow them, despite them being listed in the e820 map], i think the more robust solution is to recognize them iniounmap()as well - this way it cannot be forgotten accidentally. (and UC pages in the buddy are _hard_ to notice after the fact) There is no aliasing danger i believe: IO bars should never be marked as general RAM in the e820.
agreed, esp for .25
it's sort of a weird case of ioremap() use; I wonder if longer term we need to have a different sort of interface for this kind of use...
Quoted 3 lines · Ingo Molnar
there are many GART drivers, and the method used depends on the GART driver. The following GART drivers still use ioremap in one way or another:
There is another issue I see in the recent pageattr changes, again in the GART driver context.
GART drivers use map_page_into_agp() and unmap_page_from_agp() during runtime (and not just during load and unload of the driver module). In the recent pageattr changes, we seem to have dropped the concept of reverting to the large page(for the kernel identity mapping) while changing the attribute back to WB. In this GART driver context, over the time, potentially kernel identity mappings might be backed by small pages, if we don't revert to large page again during set_memory_wb() which gets called during unmap_page_from_agp() for a RAM page. And thus loosing the advantage of large page mapping for kernel identity mappings.
thanks,
suresh