Hi,
I've spent some time testing GPUs on aarch64 and fixing bugs.
Attached are some patches to make things more usable.
I would like to commit these or get feedback on if changes should be
implemented differently. Since I touched things in both MD code
and UVM, review would be appreciated. First line of each patch
contains a summary of what it does.
amdgpu (bonaire C1K) on AADK: works[1][2]
glamor, picom, alacritty, doomlegacy OK
radeon (bonaire) on AADK: works[1], ditto
radeon (cedar) on LX2K: works, minor rendering glitches
[1] AADK firmware doesn't POST the GPU and VGA BIOS extraction fails.
Must provide a VGA BIOS image dump manually, build a custom
EDK2 with VGA POSTing capability or figure out how to map ROM.
[2] /libdata/firmware/amdgpu is not shipped.
(Can we include these in the sets?)
Kind regards,
-Tobias
This patch set adds a new UVM freelist on aarch64 to manage pages
allocated to GPU buffers, which have a constraint on paddr < 40 bits.
Index: sys/arch/aarch64/include/vmparam.h
===
RCS file: /cvsroot/src/sys/arch/aarch64/include/vmparam.h,v
retrieving revision 1.20
diff -p -u -r1.20 vmparam.h
--- sys/arch/aarch64/include/vmparam.h 16 Apr 2023 14:01:51 - 1.20
+++ sys/arch/aarch64/include/vmparam.h 27 Dec 2023 22:07:38 -
@@ -182,8 +182,9 @@
#define VM_PHYSSEG_MAX 64 /* XXX */
#define VM_PHYSSEG_STRAT VM_PSTRAT_BSEARCH
-#define VM_NFREELIST 1
+#define VM_NFREELIST 2
#define VM_FREELIST_DEFAULT0
+#define VM_FREELIST_FIRST1T1
#elif defined(__arm__)
Index: sys/arch/evbarm/fdt/fdt_machdep.c
===
RCS file: /cvsroot/src/sys/arch/evbarm/fdt/fdt_machdep.c,v
retrieving revision 1.106
diff -p -u -r1.106 fdt_machdep.c
--- sys/arch/evbarm/fdt/fdt_machdep.c 4 Aug 2023 09:06:33 - 1.106
+++ sys/arch/evbarm/fdt/fdt_machdep.c 27 Dec 2023 22:07:38 -
@@ -194,6 +194,11 @@ fdt_add_boot_physmem(const struct fdt_me
bp->bp_start = atop(saddr);
bp->bp_pages = atop(eaddr) - bp->bp_start;
bp->bp_freelist = VM_FREELIST_DEFAULT;
+#ifdef _LP64
+ if (eaddr < (1UL<<40)) {
+ bp->bp_freelist = VM_FREELIST_FIRST1T;
+ }
+#endif
#ifdef PMAP_NEED_ALLOC_POOLPAGE
const uint64_t memory_size = *(uint64_t *)arg;
Index: sys/external/bsd/drm2/include/drm/bus_dma_hacks.h
===
RCS file: /cvsroot/src/sys/external/bsd/drm2/include/drm/bus_dma_hacks.h,v
retrieving revision 1.25
diff -p -u -r1.25 bus_dma_hacks.h
--- sys/external/bsd/drm2/include/drm/bus_dma_hacks.h 19 Jul 2022 23:19:44
- 1.25
+++ sys/external/bsd/drm2/include/drm/bus_dma_hacks.h 27 Dec 2023 22:07:38
-
@@ -78,7 +78,7 @@ BUS_MEM_TO_PHYS(bus_dma_tag_t dmat, bus_
if (dr->dr_busbase <= ba && ba - dr->dr_busbase <= dr->dr_len)
return ba - dr->dr_busbase + dr->dr_sysbase;
}
- panic("bus addr has no bus address in dma tag %p: %"PRIxPADDR, dmat,
+ panic("bus addr has no paddr in dma tag %p: %"PRIxPADDR, dmat,
ba);
}
#elif defined(__sparc__) || defined(__sparc64__)
@@ -99,6 +99,8 @@ bus_dmamem_pgfl(bus_dma_tag_t tag)
{
#if defined(__i386__) || defined(__x86_64__)
return x86_select_freelist(tag->_bounce_alloc_hi - 1);
+#elif defined(__aarch64__)
+ return VM_FREELIST_FIRST1T;
#else
return VM_FREELIST_DEFAULT;
#endif
This patch set changes the type of uvm_physseg.start_hint from u_int to u_long.
Index: sys/uvm/uvm_pglist.c
===
RCS file: /cvsroot/src/sys/uvm/uvm_pglist.c,v
retrieving revision 1.90
diff -p -u -r1.90 uvm_pglist.c
--- sys/uvm/uvm_pglist.c21 Dec 2021 08:27:49 - 1.90
+++ sys/uvm/uvm_pglist.c27 Dec 2023 22:06:10 -
@@ -112,8 +112,9 @@ static int
uvm_pglistalloc_c_ps(uvm_physseg_t psi, int num, paddr_t low, paddr_t high,
paddr_t alignment, paddr_t boundary, struct pglist *rlist)
{
- signed int candidate, limit, candidateidx, end, idx, skip;
- int pagemask;
+ long candidate, limit, candidateidx, end, idx;
+ int skip;
+ long pagemask;
bool second_pass;
#ifdef DEBUG
paddr_t idxpa, lastidxpa;
@@ -138,9 +139,9 @@ uvm_pglistalloc_c_ps(uvm_physseg_t psi,
* succeeded.
*/
alignment = atop(alignment);
- candidate = roundup2(uimax(low, uvm_physseg_get_avail_start(psi) +
+ candidate = roundup2(ulmax(low, uvm_physseg_get_avail_start(psi) +
uvm_physseg_get_start_hint(psi)), alignment);
- limit = uimin(high, uvm_physseg_get_avail_end(psi));
+ limit = ulmin(high, uvm_physseg_get_avail_end(psi));
pagemask = ~((boundary >> PAGE_SHIFT) - 1);
skip = 0;
second_pass = false;
@@ -162,8 +163,8
