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 -0000 1.20
+++ sys/arch/aarch64/include/vmparam.h 27 Dec 2023 22:07:38 -0000
@@ -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_DEFAULT 0
+#define VM_FREELIST_FIRST1T 1
#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 -0000 1.106
+++ sys/arch/evbarm/fdt/fdt_machdep.c 27 Dec 2023 22:07:38 -0000
@@ -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
-0000 1.25
+++ sys/external/bsd/drm2/include/drm/bus_dma_hacks.h 27 Dec 2023 22:07:38
-0000
@@ -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.c 21 Dec 2021 08:27:49 -0000 1.90
+++ sys/uvm/uvm_pglist.c 27 Dec 2023 22:06:10 -0000
@@ -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 @@ uvm_pglistalloc_c_ps(uvm_physseg_t psi,
* is were we started.
*/
second_pass = true;
- candidate = roundup2(uimax(low,
uvm_physseg_get_avail_start(psi)), alignment);
- limit = uimin(limit, uvm_physseg_get_avail_start(psi) +
+ candidate = roundup2(ulmax(low,
uvm_physseg_get_avail_start(psi)), alignment);
+ limit = ulmin(limit, uvm_physseg_get_avail_start(psi) +
uvm_physseg_get_start_hint(psi));
skip = 0;
continue;
@@ -200,7 +201,7 @@ uvm_pglistalloc_c_ps(uvm_physseg_t psi,
* Found a suitable starting page. See if the range is free.
*/
#ifdef PGALLOC_VERBOSE
- printf("%s: psi=%d candidate=%#x end=%#x skip=%#x,
align=%#"PRIxPADDR,
+ printf("%s: psi=%d candidate=%#lx end=%#lx skip=%#x,
align=%#"PRIxPADDR,
__func__, psi, candidateidx, end, skip, alignment);
#endif
/*
@@ -283,7 +284,7 @@ uvm_pglistalloc_c_ps(uvm_physseg_t psi,
uvm_physseg_get_avail_start(psi));
KASSERTMSG(uvm_physseg_get_start_hint(psi) <=
uvm_physseg_get_avail_end(psi) - uvm_physseg_get_avail_start(psi),
- "%x %u (%#x) <= %#"PRIxPADDR" - %#"PRIxPADDR" (%#"PRIxPADDR")",
+ "%lx %lu (%#lx) <= %#"PRIxPADDR" - %#"PRIxPADDR" (%#"PRIxPADDR")",
candidate + num,
uvm_physseg_get_start_hint(psi), uvm_physseg_get_start_hint(psi),
uvm_physseg_get_avail_end(psi), uvm_physseg_get_avail_start(psi),
@@ -523,7 +524,8 @@ static int
uvm_pglistalloc_s_ps(uvm_physseg_t psi, int num, paddr_t low, paddr_t high,
struct pglist *rlist)
{
- int todo, limit, candidate;
+ int todo;
+ long limit, candidate;
struct vm_page *pg;
bool second_pass;
#ifdef PGALLOC_VERBOSE
@@ -546,9 +548,9 @@ uvm_pglistalloc_s_ps(uvm_physseg_t psi,
return -1;
todo = num;
- candidate = uimax(low, uvm_physseg_get_avail_start(psi) +
+ candidate = ulmax(low, uvm_physseg_get_avail_start(psi) +
uvm_physseg_get_start_hint(psi));
- limit = uimin(high, uvm_physseg_get_avail_end(psi));
+ limit = ulmin(high, uvm_physseg_get_avail_end(psi));
pg = uvm_physseg_get_pg(psi, candidate - uvm_physseg_get_start(psi));
second_pass = false;
@@ -560,8 +562,8 @@ again:
break;
}
second_pass = true;
- candidate = uimax(low,
uvm_physseg_get_avail_start(psi));
- limit = uimin(limit, uvm_physseg_get_avail_start(psi) +
+ candidate = ulmax(low,
uvm_physseg_get_avail_start(psi));
+ limit = ulmin(limit, uvm_physseg_get_avail_start(psi) +
uvm_physseg_get_start_hint(psi));
pg = uvm_physseg_get_pg(psi, candidate -
uvm_physseg_get_start(psi));
goto again;
@@ -571,10 +573,10 @@ again:
paddr_t cidx = 0;
const uvm_physseg_t bank = uvm_physseg_find(candidate,
&cidx);
KDASSERTMSG(bank == psi,
- "uvm_physseg_find(%#x) (%"PRIxPHYSSEG ") != psi
%"PRIxPHYSSEG,
+ "uvm_physseg_find(%#lx) (%"PRIxPHYSSEG ") != psi
%"PRIxPHYSSEG,
candidate, bank, psi);
KDASSERTMSG(cidx == candidate -
uvm_physseg_get_start(psi),
- "uvm_physseg_find(%#x): %#"PRIxPADDR" != off
%"PRIxPADDR,
+ "uvm_physseg_find(%#lx): %#"PRIxPADDR" != off
%"PRIxPADDR,
candidate, cidx, candidate -
uvm_physseg_get_start(psi));
}
#endif
@@ -594,7 +596,7 @@ again:
uvm_physseg_set_start_hint(psi, candidate + 1 -
uvm_physseg_get_avail_start(psi));
KASSERTMSG(uvm_physseg_get_start_hint(psi) <=
uvm_physseg_get_avail_end(psi) -
uvm_physseg_get_avail_start(psi),
- "%#x %u (%#x) <= %#"PRIxPADDR" - %#"PRIxPADDR" (%#"PRIxPADDR")",
+ "%#lx %lu (%#lx) <= %#"PRIxPADDR" - %#"PRIxPADDR" (%#"PRIxPADDR")",
candidate + 1,
uvm_physseg_get_start_hint(psi),
uvm_physseg_get_start_hint(psi),
Index: sys/uvm/uvm_physseg.c
===================================================================
RCS file: /cvsroot/src/sys/uvm/uvm_physseg.c,v
retrieving revision 1.19
diff -p -u -r1.19 uvm_physseg.c
--- sys/uvm/uvm_physseg.c 23 Sep 2023 18:20:20 -0000 1.19
+++ sys/uvm/uvm_physseg.c 27 Dec 2023 22:06:10 -0000
@@ -100,7 +100,7 @@ struct uvm_physseg {
paddr_t avail_end; /* (PF# of last free page in segment)
+1 */
struct extent *ext; /* extent(9) structure to manage pgs[]
*/
int free_list; /* which free list they belong on */
- u_int start_hint; /* start looking for free pages here */
+ u_long start_hint; /* start looking for free pages here */
#ifdef __HAVE_PMAP_PHYSSEG
struct pmap_physseg pmseg; /* pmap specific (MD) data */
#endif
@@ -1108,7 +1108,7 @@ uvm_physseg_get_free_list(uvm_physseg_t
return HANDLE_TO_PHYSSEG_NODE(upm)->free_list;
}
-u_int
+u_long
uvm_physseg_get_start_hint(uvm_physseg_t upm)
{
KASSERT(uvm_physseg_valid_p(upm));
@@ -1116,7 +1116,7 @@ uvm_physseg_get_start_hint(uvm_physseg_t
}
bool
-uvm_physseg_set_start_hint(uvm_physseg_t upm, u_int start_hint)
+uvm_physseg_set_start_hint(uvm_physseg_t upm, u_long start_hint)
{
if (uvm_physseg_valid_p(upm) == false)
return false;
Index: sys/uvm/uvm_physseg.h
===================================================================
RCS file: /cvsroot/src/sys/uvm/uvm_physseg.h,v
retrieving revision 1.8
diff -p -u -r1.8 uvm_physseg.h
--- sys/uvm/uvm_physseg.h 2 Jan 2017 20:08:32 -0000 1.8
+++ sys/uvm/uvm_physseg.h 27 Dec 2023 22:06:10 -0000
@@ -106,8 +106,8 @@ struct pmap_physseg * uvm_physseg_get_pm
#endif
int uvm_physseg_get_free_list(uvm_physseg_t);
-u_int uvm_physseg_get_start_hint(uvm_physseg_t);
-bool uvm_physseg_set_start_hint(uvm_physseg_t, u_int);
+u_long uvm_physseg_get_start_hint(uvm_physseg_t);
+bool uvm_physseg_set_start_hint(uvm_physseg_t, u_long);
/*
* Functions to help walk the list of segments.
This patch works around rendering corruption on aarch64.
Index: external/mit/MesaLib/dist/src/gallium/drivers/radeonsi/si_buffer.c
===================================================================
RCS file:
/cvsroot/xsrc/external/mit/MesaLib/dist/src/gallium/drivers/radeonsi/si_buffer.c,v
retrieving revision 1.1.1.3
diff -p -u -r1.1.1.3 si_buffer.c
--- external/mit/MesaLib/dist/src/gallium/drivers/radeonsi/si_buffer.c 9 May
2022 01:23:33 -0000 1.1.1.3
+++ external/mit/MesaLib/dist/src/gallium/drivers/radeonsi/si_buffer.c 27 Dec
2023 23:08:23 -0000
@@ -100,6 +100,10 @@ void si_init_resource_fields(struct si_s
*/
if (!sscreen->info.kernel_flushes_hdp_before_ib ||
!sscreen->info.is_amdgpu)
res->domains = RADEON_DOMAIN_GTT;
+#if defined(PIPE_ARCH_AARCH64)
+ /* https://gitlab.freedesktop.org/mesa/mesa/-/issues/3274 */
+ res->domains = RADEON_DOMAIN_GTT;
+#endif
}
/* Tiled textures are unmappable. Always put them in VRAM. */
