2.6.14-rc4 AGP performance fixes
PATCH 1/1 2.6.14-rc4 AGP performance fixes
AGP allocation/deallocation is suffering major performance issues due to the nature of global_flush_tlb() being called on every change_page_attr() call.
For small allocations this isn't really seen, but when you start allocating 50000 pages of AGP space, for say, texture memory, then things can take seconds to complete.
In some cases the situation is doubled or even quadrupled in the time due to SMP, or a deallocation, then a new reallocation. I've had a case of upto 20 seconds wait time to deallocate and reallocate AGP space.
This patch fixes the problem by making it the caller's responsibility to call global_flush_tlb(), and so removes it from every instance of mapping a page into AGP space until the time that all change_page_attr() changes are done.
Patch is against 2.6.14-rc4.
Alan.
drivers/char/agp/backend.c +7 -2
@@ -147,6 +147,7 @@ printk(KERN_ERR PFX "unable to get memory for scratch page.\n");return -ENOMEM ; } +global_flush_tlb (); bridge->scratch_page_real =virt_to_gart (addr); bridge->scratch_page =
@@ -187,9 +188,11 @@return 0 ; err_out: -if (bridge->driver->needs_scratch_page) +if (bridge->driver->needs_scratch_page) { bridge->driver->agp_destroy_page (gart_to_virt (bridge->scratch_page_real)); +global_flush_tlb (); + }if (got_gatt) bridge->driver->free_gatt_table (bridge);if (got_keylist) {
@@ -211,9 +214,11 @@ bridge->key_list =NULL ;if (bridge->driver->agp_destroy_page && - bridge->driver->needs_scratch_page) + bridge->driver->needs_scratch_page) { bridge->driver->agp_destroy_page (gart_to_virt (bridge->scratch_page_real)); +global_flush_tlb (); + } }/* When we remove the global variable agp_bridge from all drivers
drivers/char/agp/generic.c +8 -3
@@ -57,7 +57,8 @@ {int i; i =change_page_attr (page,1 ,PAGE_KERNEL_NOCACHE ); -global_flush_tlb (); +/* Caller's responsibility to call global_flush_tlb() for +* performance reasons */ return i; }EXPORT_SYMBOL_GPL (map_page_into_agp);
@@ -66,7 +67,8 @@ {int i; i =change_page_attr (page,1 ,PAGE_KERNEL ); -global_flush_tlb (); +/* Caller's responsibility to call global_flush_tlb() for +* performance reasons */ return i; }EXPORT_SYMBOL_GPL (unmap_page_from_agp);
@@ -155,6 +157,7 @@for (i =0 ; i < curr->page_count; i++) { curr->bridge->driver->agp_destroy_page (gart_to_virt (curr->memory[i])); } +global_flush_tlb (); }agp_free_key (curr->key);vfree (curr->memory);
@@ -212,7 +215,9 @@ new->memory[i] =virt_to_gart (addr); new->page_count++; } - new->bridge = bridge; +global_flush_tlb (); + + new->bridge = bridge;flush_agp_mappings ();
drivers/char/agp/i460-agp.c +6 -3
@@ -514,9 +514,10 @@ {void *page; -if (I460_IO_PAGE_SHIFT <=PAGE_SHIFT ) +if (I460_IO_PAGE_SHIFT <=PAGE_SHIFT ) { page =agp_generic_alloc_page (agp_bridge); -else +global_flush_tlb (); + }else /* Returning NULL would cause problems */ /* AK: really dubious code. */ page = (void *)~0UL ;
@@ -525,8 +526,10 @@static void i460_destroy_page (void *page) { -if (I460_IO_PAGE_SHIFT <=PAGE_SHIFT ) +if (I460_IO_PAGE_SHIFT <=PAGE_SHIFT ) {agp_generic_destroy_page (page); +global_flush_tlb (); + } }#endif /* I460_LARGE_IO_PAGES */
drivers/char/agp/intel-agp.c +4 -1
@@ -270,6 +270,7 @@switch (pg_count) {case 1 : addr = agp_bridge->driver->agp_alloc_page (agp_bridge); +global_flush_tlb ();break ;case 4 :/* kludge to get 4 physical pages for ARGB cursor */
@@ -330,9 +331,11 @@if (curr->type ==AGP_PHYS_MEMORY ) {if (curr->page_count ==4 )i8xx_destroy_pages (gart_to_virt (curr->memory[0 ])); -else +else { agp_bridge->driver->agp_destroy_page (gart_to_virt (curr->memory[0 ])); +global_flush_tlb (); + }vfree (curr->memory); }kfree (curr);
Quoted 10 lines · Alan Hourihane
AGP allocation/deallocation is suffering major performance issues due to the nature of
global_flush_tlb()being called on everychange_page_attr()call.For small allocations this isn't really seen, but when you start allocating 50000 pages of AGP space, for say, texture memory, then things can take seconds to complete.
In some cases the situation is doubled or even quadrupled in the time due to SMP, or a deallocation, then a new reallocation. I've had a case of upto 20 seconds wait time to deallocate and reallocate AGP space.
Yikes.
Quoted 4 lines
This patch fixes the problem by making it the caller's responsibility to call
global_flush_tlb(), and so removes it from every instance of mapping a page into AGP space until the time that allchange_page_attr()changes are done.
I like the idea of minimising the flushes where we can, however the idea of having stale entries in the TLB's, even for a short time gives me the jibblies a little. If this does cause any problems, they're likely to be of a nature that would be incredibly difficult to track down.
(We've been bitten on more than one occasion due to missing flushes in this driver). Given the wide variety of hardware this driver supports, I'd like this to sit in -mm for a while, just to be on the safe side.
If Andrew picks that up now, that gives us a while for testing before 2.6.15 closes for features.
Dave