]> git.ipfire.org Git - thirdparty/linux.git/commitdiff
drm/amdkfd: Handle deallocated VPGRs in gfx11+ trap handler
authorJay Cornwall <jay.cornwall@amd.com>
Tue, 28 May 2024 20:55:56 +0000 (15:55 -0500)
committerAlex Deucher <alexander.deucher@amd.com>
Wed, 5 Jun 2024 15:05:57 +0000 (11:05 -0400)
A wavefront may deallocate its VGPRs at the end of a program while
waiting for memory transactions to complete. If it subsequently
receives a context save exception it will be unable to save,
since this requires VGPRs. In this case the trap handler should
terminate the wavefront.

Fixes intermittent VM faults under context switching load.

V2: Use S_ENDPGM instead of S_ENDPGM_SAVED for performance counters

Signed-off-by: Jay Cornwall <jay.cornwall@amd.com>
Reviewed-by: Lancelot Six <lancelot.six@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
drivers/gpu/drm/amd/amdkfd/cwsr_trap_handler.h
drivers/gpu/drm/amd/amdkfd/cwsr_trap_handler_gfx10.asm

index 85a41e121cce1b26b4ad14588a6e8769acbc712d..665122d1bbbdf92be1e87116259369a2b24ce553 100644 (file)
@@ -2705,7 +2705,7 @@ static const uint32_t cwsr_trap_gfx10_hex[] = {
 };
 
 static const uint32_t cwsr_trap_gfx11_hex[] = {
-       0xbfa00001, 0xbfa00224,
+       0xbfa00001, 0xbfa00227,
        0xb0804006, 0xb8f8f802,
        0x9178ff78, 0x00020006,
        0xb8fbf803, 0xbf0d9e6d,
@@ -2750,399 +2750,400 @@ static const uint32_t cwsr_trap_gfx11_hex[] = {
        0x8b6dff6d, 0x0000ffff,
        0x8bfe7e7e, 0x8bea6a6a,
        0xb978f802, 0xbe804a6c,
-       0x8b6dff6d, 0x0000ffff,
-       0xbefa0080, 0xb97a0283,
-       0xbeee007e, 0xbeef007f,
-       0xbefe0180, 0xbefe4d84,
-       0xbf890000, 0x8b7aff7f,
-       0x04000000, 0x847a857a,
-       0x8c6d7a6d, 0xbefa007e,
-       0x8b7bff7f, 0x0000ffff,
-       0xbefe00c1, 0xbeff00c1,
-       0xdca6c000, 0x007a0000,
-       0x7e000280, 0xbefe007a,
-       0xbeff007b, 0xb8fb02dc,
-       0x847b997b, 0xb8fa3b05,
-       0x807a817a, 0xbf0d997b,
-       0xbfa20002, 0x847a897a,
-       0xbfa00001, 0x847a8a7a,
-       0xb8fb1e06, 0x847b8a7b,
-       0x807a7b7a, 0x8b7bff7f,
-       0x0000ffff, 0x807aff7a,
-       0x00000200, 0x807a7e7a,
-       0x827b807b, 0xd7610000,
-       0x00010870, 0xd7610000,
-       0x00010a71, 0xd7610000,
-       0x00010c72, 0xd7610000,
-       0x00010e73, 0xd7610000,
-       0x00011074, 0xd7610000,
-       0x00011275, 0xd7610000,
-       0x00011476, 0xd7610000,
-       0x00011677, 0xd7610000,
-       0x00011a79, 0xd7610000,
-       0x00011c7e, 0xd7610000,
-       0x00011e7f, 0xbefe00ff,
-       0x00003fff, 0xbeff0080,
-       0xdca6c040, 0x007a0000,
-       0xd760007a, 0x00011d00,
-       0xd760007b, 0x00011f00,
+       0xbf0d9878, 0xbfa10001,
+       0xbfb00000, 0x8b6dff6d,
+       0x0000ffff, 0xbefa0080,
+       0xb97a0283, 0xbeee007e,
+       0xbeef007f, 0xbefe0180,
+       0xbefe4d84, 0xbf890000,
+       0x8b7aff7f, 0x04000000,
+       0x847a857a, 0x8c6d7a6d,
+       0xbefa007e, 0x8b7bff7f,
+       0x0000ffff, 0xbefe00c1,
+       0xbeff00c1, 0xdca6c000,
+       0x007a0000, 0x7e000280,
        0xbefe007a, 0xbeff007b,
-       0xbef4007e, 0x8b75ff7f,
-       0x0000ffff, 0x8c75ff75,
-       0x00040000, 0xbef60080,
-       0xbef700ff, 0x10807fac,
-       0xbef1007d, 0xbef00080,
-       0xb8f302dc, 0x84739973,
-       0xbefe00c1, 0x857d9973,
-       0x8b7d817d, 0xbf06817d,
-       0xbfa20002, 0xbeff0080,
-       0xbfa00002, 0xbeff00c1,
-       0xbfa00009, 0xbef600ff,
-       0x01000000, 0xe0685080,
-       0x701d0100, 0xe0685100,
-       0x701d0200, 0xe0685180,
-       0x701d0300, 0xbfa00008,
+       0xb8fb02dc, 0x847b997b,
+       0xb8fa3b05, 0x807a817a,
+       0xbf0d997b, 0xbfa20002,
+       0x847a897a, 0xbfa00001,
+       0x847a8a7a, 0xb8fb1e06,
+       0x847b8a7b, 0x807a7b7a,
+       0x8b7bff7f, 0x0000ffff,
+       0x807aff7a, 0x00000200,
+       0x807a7e7a, 0x827b807b,
+       0xd7610000, 0x00010870,
+       0xd7610000, 0x00010a71,
+       0xd7610000, 0x00010c72,
+       0xd7610000, 0x00010e73,
+       0xd7610000, 0x00011074,
+       0xd7610000, 0x00011275,
+       0xd7610000, 0x00011476,
+       0xd7610000, 0x00011677,
+       0xd7610000, 0x00011a79,
+       0xd7610000, 0x00011c7e,
+       0xd7610000, 0x00011e7f,
+       0xbefe00ff, 0x00003fff,
+       0xbeff0080, 0xdca6c040,
+       0x007a0000, 0xd760007a,
+       0x00011d00, 0xd760007b,
+       0x00011f00, 0xbefe007a,
+       0xbeff007b, 0xbef4007e,
+       0x8b75ff7f, 0x0000ffff,
+       0x8c75ff75, 0x00040000,
+       0xbef60080, 0xbef700ff,
+       0x10807fac, 0xbef1007d,
+       0xbef00080, 0xb8f302dc,
+       0x84739973, 0xbefe00c1,
+       0x857d9973, 0x8b7d817d,
+       0xbf06817d, 0xbfa20002,
+       0xbeff0080, 0xbfa00002,
+       0xbeff00c1, 0xbfa00009,
        0xbef600ff, 0x01000000,
-       0xe0685100, 0x701d0100,
-       0xe0685200, 0x701d0200,
-       0xe0685300, 0x701d0300,
+       0xe0685080, 0x701d0100,
+       0xe0685100, 0x701d0200,
+       0xe0685180, 0x701d0300,
+       0xbfa00008, 0xbef600ff,
+       0x01000000, 0xe0685100,
+       0x701d0100, 0xe0685200,
+       0x701d0200, 0xe0685300,
+       0x701d0300, 0xb8f03b05,
+       0x80708170, 0xbf0d9973,
+       0xbfa20002, 0x84708970,
+       0xbfa00001, 0x84708a70,
+       0xb8fa1e06, 0x847a8a7a,
+       0x80707a70, 0x8070ff70,
+       0x00000200, 0xbef600ff,
+       0x01000000, 0x7e000280,
+       0x7e020280, 0x7e040280,
+       0xbefd0080, 0xd7610002,
+       0x0000fa71, 0x807d817d,
+       0xd7610002, 0x0000fa6c,
+       0x807d817d, 0x917aff6d,
+       0x80000000, 0xd7610002,
+       0x0000fa7a, 0x807d817d,
+       0xd7610002, 0x0000fa6e,
+       0x807d817d, 0xd7610002,
+       0x0000fa6f, 0x807d817d,
+       0xd7610002, 0x0000fa78,
+       0x807d817d, 0xb8faf803,
+       0xd7610002, 0x0000fa7a,
+       0x807d817d, 0xd7610002,
+       0x0000fa7b, 0x807d817d,
+       0xb8f1f801, 0xd7610002,
+       0x0000fa71, 0x807d817d,
+       0xb8f1f814, 0xd7610002,
+       0x0000fa71, 0x807d817d,
+       0xb8f1f815, 0xd7610002,
+       0x0000fa71, 0x807d817d,
+       0xbefe00ff, 0x0000ffff,
+       0xbeff0080, 0xe0685000,
+       0x701d0200, 0xbefe00c1,
        0xb8f03b05, 0x80708170,
        0xbf0d9973, 0xbfa20002,
        0x84708970, 0xbfa00001,
        0x84708a70, 0xb8fa1e06,
        0x847a8a7a, 0x80707a70,
-       0x8070ff70, 0x00000200,
        0xbef600ff, 0x01000000,
-       0x7e000280, 0x7e020280,
-       0x7e040280, 0xbefd0080,
-       0xd7610002, 0x0000fa71,
-       0x807d817d, 0xd7610002,
-       0x0000fa6c, 0x807d817d,
-       0x917aff6d, 0x80000000,
-       0xd7610002, 0x0000fa7a,
-       0x807d817d, 0xd7610002,
-       0x0000fa6e, 0x807d817d,
-       0xd7610002, 0x0000fa6f,
-       0x807d817d, 0xd7610002,
-       0x0000fa78, 0x807d817d,
-       0xb8faf803, 0xd7610002,
-       0x0000fa7a, 0x807d817d,
-       0xd7610002, 0x0000fa7b,
-       0x807d817d, 0xb8f1f801,
-       0xd7610002, 0x0000fa71,
-       0x807d817d, 0xb8f1f814,
-       0xd7610002, 0x0000fa71,
-       0x807d817d, 0xb8f1f815,
-       0xd7610002, 0x0000fa71,
-       0x807d817d, 0xbefe00ff,
-       0x0000ffff, 0xbeff0080,
-       0xe0685000, 0x701d0200,
-       0xbefe00c1, 0xb8f03b05,
-       0x80708170, 0xbf0d9973,
-       0xbfa20002, 0x84708970,
-       0xbfa00001, 0x84708a70,
-       0xb8fa1e06, 0x847a8a7a,
-       0x80707a70, 0xbef600ff,
-       0x01000000, 0xbef90080,
-       0xbefd0080, 0xbf800000,
-       0xbe804100, 0xbe824102,
-       0xbe844104, 0xbe864106,
-       0xbe884108, 0xbe8a410a,
-       0xbe8c410c, 0xbe8e410e,
-       0xd7610002, 0x0000f200,
-       0x80798179, 0xd7610002,
-       0x0000f201, 0x80798179,
-       0xd7610002, 0x0000f202,
-       0x80798179, 0xd7610002,
-       0x0000f203, 0x80798179,
-       0xd7610002, 0x0000f204,
+       0xbef90080, 0xbefd0080,
+       0xbf800000, 0xbe804100,
+       0xbe824102, 0xbe844104,
+       0xbe864106, 0xbe884108,
+       0xbe8a410a, 0xbe8c410c,
+       0xbe8e410e, 0xd7610002,
+       0x0000f200, 0x80798179,
+       0xd7610002, 0x0000f201,
        0x80798179, 0xd7610002,
-       0x0000f205, 0x80798179,
-       0xd7610002, 0x0000f206,
+       0x0000f202, 0x80798179,
+       0xd7610002, 0x0000f203,
        0x80798179, 0xd7610002,
-       0x0000f207, 0x80798179,
-       0xd7610002, 0x0000f208,
+       0x0000f204, 0x80798179,
+       0xd7610002, 0x0000f205,
        0x80798179, 0xd7610002,
-       0x0000f209, 0x80798179,
-       0xd7610002, 0x0000f20a,
+       0x0000f206, 0x80798179,
+       0xd7610002, 0x0000f207,
        0x80798179, 0xd7610002,
-       0x0000f20b, 0x80798179,
-       0xd7610002, 0x0000f20c,
+       0x0000f208, 0x80798179,
+       0xd7610002, 0x0000f209,
        0x80798179, 0xd7610002,
-       0x0000f20d, 0x80798179,
-       0xd7610002, 0x0000f20e,
+       0x0000f20a, 0x80798179,
+       0xd7610002, 0x0000f20b,
        0x80798179, 0xd7610002,
-       0x0000f20f, 0x80798179,
-       0xbf06a079, 0xbfa10006,
-       0xe0685000, 0x701d0200,
-       0x8070ff70, 0x00000080,
-       0xbef90080, 0x7e040280,
-       0x807d907d, 0xbf0aff7d,
-       0x00000060, 0xbfa2ffbc,
-       0xbe804100, 0xbe824102,
-       0xbe844104, 0xbe864106,
-       0xbe884108, 0xbe8a410a,
-       0xd7610002, 0x0000f200,
+       0x0000f20c, 0x80798179,
+       0xd7610002, 0x0000f20d,
        0x80798179, 0xd7610002,
-       0x0000f201, 0x80798179,
-       0xd7610002, 0x0000f202,
+       0x0000f20e, 0x80798179,
+       0xd7610002, 0x0000f20f,
+       0x80798179, 0xbf06a079,
+       0xbfa10006, 0xe0685000,
+       0x701d0200, 0x8070ff70,
+       0x00000080, 0xbef90080,
+       0x7e040280, 0x807d907d,
+       0xbf0aff7d, 0x00000060,
+       0xbfa2ffbc, 0xbe804100,
+       0xbe824102, 0xbe844104,
+       0xbe864106, 0xbe884108,
+       0xbe8a410a, 0xd7610002,
+       0x0000f200, 0x80798179,
+       0xd7610002, 0x0000f201,
        0x80798179, 0xd7610002,
-       0x0000f203, 0x80798179,
-       0xd7610002, 0x0000f204,
+       0x0000f202, 0x80798179,
+       0xd7610002, 0x0000f203,
        0x80798179, 0xd7610002,
-       0x0000f205, 0x80798179,
-       0xd7610002, 0x0000f206,
+       0x0000f204, 0x80798179,
+       0xd7610002, 0x0000f205,
        0x80798179, 0xd7610002,
-       0x0000f207, 0x80798179,
-       0xd7610002, 0x0000f208,
+       0x0000f206, 0x80798179,
+       0xd7610002, 0x0000f207,
        0x80798179, 0xd7610002,
-       0x0000f209, 0x80798179,
-       0xd7610002, 0x0000f20a,
+       0x0000f208, 0x80798179,
+       0xd7610002, 0x0000f209,
        0x80798179, 0xd7610002,
-       0x0000f20b, 0x80798179,
-       0xe0685000, 0x701d0200,
-       0xbefe00c1, 0x857d9973,
-       0x8b7d817d, 0xbf06817d,
-       0xbfa20002, 0xbeff0080,
-       0xbfa00001, 0xbeff00c1,
-       0xb8fb4306, 0x8b7bc17b,
-       0xbfa10043, 0xbfbd0000,
-       0x8b7aff6d, 0x80000000,
-       0xbfa1003f, 0x847b887b,
-       0xbef6007b, 0xb8f03b05,
-       0x80708170, 0xbf0d9973,
-       0xbfa20002, 0x84708970,
-       0xbfa00001, 0x84708a70,
-       0xb8fa1e06, 0x847a8a7a,
-       0x80707a70, 0x8070ff70,
-       0x00000200, 0x8070ff70,
-       0x00000080, 0xbef600ff,
-       0x01000000, 0xd71f0000,
-       0x000100c1, 0xd7200000,
-       0x000200c1, 0x16000084,
+       0x0000f20a, 0x80798179,
+       0xd7610002, 0x0000f20b,
+       0x80798179, 0xe0685000,
+       0x701d0200, 0xbefe00c1,
        0x857d9973, 0x8b7d817d,
-       0xbf06817d, 0xbefd0080,
-       0xbfa20012, 0xbe8300ff,
-       0x00000080, 0xbf800000,
+       0xbf06817d, 0xbfa20002,
+       0xbeff0080, 0xbfa00001,
+       0xbeff00c1, 0xb8fb4306,
+       0x8b7bc17b, 0xbfa10043,
+       0xbfbd0000, 0x8b7aff6d,
+       0x80000000, 0xbfa1003f,
+       0x847b887b, 0xbef6007b,
+       0xb8f03b05, 0x80708170,
+       0xbf0d9973, 0xbfa20002,
+       0x84708970, 0xbfa00001,
+       0x84708a70, 0xb8fa1e06,
+       0x847a8a7a, 0x80707a70,
+       0x8070ff70, 0x00000200,
+       0x8070ff70, 0x00000080,
+       0xbef600ff, 0x01000000,
+       0xd71f0000, 0x000100c1,
+       0xd7200000, 0x000200c1,
+       0x16000084, 0x857d9973,
+       0x8b7d817d, 0xbf06817d,
+       0xbefd0080, 0xbfa20012,
+       0xbe8300ff, 0x00000080,
        0xbf800000, 0xbf800000,
-       0xd8d80000, 0x01000000,
-       0xbf890000, 0xe0685000,
-       0x701d0100, 0x807d037d,
-       0x80700370, 0xd5250000,
-       0x0001ff00, 0x00000080,
-       0xbf0a7b7d, 0xbfa2fff4,
-       0xbfa00011, 0xbe8300ff,
-       0x00000100, 0xbf800000,
+       0xbf800000, 0xd8d80000,
+       0x01000000, 0xbf890000,
+       0xe0685000, 0x701d0100,
+       0x807d037d, 0x80700370,
+       0xd5250000, 0x0001ff00,
+       0x00000080, 0xbf0a7b7d,
+       0xbfa2fff4, 0xbfa00011,
+       0xbe8300ff, 0x00000100,
        0xbf800000, 0xbf800000,
-       0xd8d80000, 0x01000000,
-       0xbf890000, 0xe0685000,
-       0x701d0100, 0x807d037d,
-       0x80700370, 0xd5250000,
-       0x0001ff00, 0x00000100,
-       0xbf0a7b7d, 0xbfa2fff4,
-       0xbefe00c1, 0x857d9973,
-       0x8b7d817d, 0xbf06817d,
-       0xbfa20004, 0xbef000ff,
-       0x00000200, 0xbeff0080,
-       0xbfa00003, 0xbef000ff,
-       0x00000400, 0xbeff00c1,
-       0xb8fb3b05, 0x807b817b,
-       0x847b827b, 0x857d9973,
-       0x8b7d817d, 0xbf06817d,
-       0xbfa20017, 0xbef600ff,
-       0x01000000, 0xbefd0084,
-       0xbf0a7b7d, 0xbfa10037,
-       0x7e008700, 0x7e028701,
-       0x7e048702, 0x7e068703,
-       0xe0685000, 0x701d0000,
-       0xe0685080, 0x701d0100,
-       0xe0685100, 0x701d0200,
-       0xe0685180, 0x701d0300,
-       0x807d847d, 0x8070ff70,
-       0x00000200, 0xbf0a7b7d,
-       0xbfa2ffef, 0xbfa00025,
+       0xbf800000, 0xd8d80000,
+       0x01000000, 0xbf890000,
+       0xe0685000, 0x701d0100,
+       0x807d037d, 0x80700370,
+       0xd5250000, 0x0001ff00,
+       0x00000100, 0xbf0a7b7d,
+       0xbfa2fff4, 0xbefe00c1,
+       0x857d9973, 0x8b7d817d,
+       0xbf06817d, 0xbfa20004,
+       0xbef000ff, 0x00000200,
+       0xbeff0080, 0xbfa00003,
+       0xbef000ff, 0x00000400,
+       0xbeff00c1, 0xb8fb3b05,
+       0x807b817b, 0x847b827b,
+       0x857d9973, 0x8b7d817d,
+       0xbf06817d, 0xbfa20017,
        0xbef600ff, 0x01000000,
        0xbefd0084, 0xbf0a7b7d,
-       0xbfa10011, 0x7e008700,
+       0xbfa10037, 0x7e008700,
        0x7e028701, 0x7e048702,
        0x7e068703, 0xe0685000,
-       0x701d0000, 0xe0685100,
-       0x701d0100, 0xe0685200,
-       0x701d0200, 0xe0685300,
+       0x701d0000, 0xe0685080,
+       0x701d0100, 0xe0685100,
+       0x701d0200, 0xe0685180,
        0x701d0300, 0x807d847d,
-       0x8070ff70, 0x00000400,
+       0x8070ff70, 0x00000200,
        0xbf0a7b7d, 0xbfa2ffef,
-       0xb8fb1e06, 0x8b7bc17b,
-       0xbfa1000c, 0x847b837b,
-       0x807b7d7b, 0xbefe00c1,
-       0xbeff0080, 0x7e008700,
+       0xbfa00025, 0xbef600ff,
+       0x01000000, 0xbefd0084,
+       0xbf0a7b7d, 0xbfa10011,
+       0x7e008700, 0x7e028701,
+       0x7e048702, 0x7e068703,
        0xe0685000, 0x701d0000,
-       0x807d817d, 0x8070ff70,
-       0x00000080, 0xbf0a7b7d,
-       0xbfa2fff8, 0xbfa0013f,
-       0xbef4007e, 0x8b75ff7f,
-       0x0000ffff, 0x8c75ff75,
-       0x00040000, 0xbef60080,
-       0xbef700ff, 0x10807fac,
-       0xb8f202dc, 0x84729972,
-       0x8b6eff7f, 0x04000000,
-       0xbfa10039, 0xbefe00c1,
-       0x857d9972, 0x8b7d817d,
-       0xbf06817d, 0xbfa20002,
-       0xbeff0080, 0xbfa00001,
-       0xbeff00c1, 0xb8ef4306,
-       0x8b6fc16f, 0xbfa1002e,
-       0x846f886f, 0xbef6006f,
-       0xb8f83b05, 0x80788178,
-       0xbf0d9972, 0xbfa20002,
-       0x84788978, 0xbfa00001,
-       0x84788a78, 0xb8ee1e06,
-       0x846e8a6e, 0x80786e78,
+       0xe0685100, 0x701d0100,
+       0xe0685200, 0x701d0200,
+       0xe0685300, 0x701d0300,
+       0x807d847d, 0x8070ff70,
+       0x00000400, 0xbf0a7b7d,
+       0xbfa2ffef, 0xb8fb1e06,
+       0x8b7bc17b, 0xbfa1000c,
+       0x847b837b, 0x807b7d7b,
+       0xbefe00c1, 0xbeff0080,
+       0x7e008700, 0xe0685000,
+       0x701d0000, 0x807d817d,
+       0x8070ff70, 0x00000080,
+       0xbf0a7b7d, 0xbfa2fff8,
+       0xbfa0013f, 0xbef4007e,
+       0x8b75ff7f, 0x0000ffff,
+       0x8c75ff75, 0x00040000,
+       0xbef60080, 0xbef700ff,
+       0x10807fac, 0xb8f202dc,
+       0x84729972, 0x8b6eff7f,
+       0x04000000, 0xbfa10039,
+       0xbefe00c1, 0x857d9972,
+       0x8b7d817d, 0xbf06817d,
+       0xbfa20002, 0xbeff0080,
+       0xbfa00001, 0xbeff00c1,
+       0xb8ef4306, 0x8b6fc16f,
+       0xbfa1002e, 0x846f886f,
+       0xbef6006f, 0xb8f83b05,
+       0x80788178, 0xbf0d9972,
+       0xbfa20002, 0x84788978,
+       0xbfa00001, 0x84788a78,
+       0xb8ee1e06, 0x846e8a6e,
+       0x80786e78, 0x8078ff78,
+       0x00000200, 0x8078ff78,
+       0x00000080, 0xbef600ff,
+       0x01000000, 0x857d9972,
+       0x8b7d817d, 0xbf06817d,
+       0xbefd0080, 0xbfa2000c,
+       0xe0500000, 0x781d0000,
+       0xbf890000, 0xdac00000,
+       0x00000000, 0x807dff7d,
+       0x00000080, 0x8078ff78,
+       0x00000080, 0xbf0a6f7d,
+       0xbfa2fff5, 0xbfa0000b,
+       0xe0500000, 0x781d0000,
+       0xbf890000, 0xdac00000,
+       0x00000000, 0x807dff7d,
+       0x00000100, 0x8078ff78,
+       0x00000100, 0xbf0a6f7d,
+       0xbfa2fff5, 0xbef80080,
+       0xbefe00c1, 0x857d9972,
+       0x8b7d817d, 0xbf06817d,
+       0xbfa20002, 0xbeff0080,
+       0xbfa00001, 0xbeff00c1,
+       0xb8ef3b05, 0x806f816f,
+       0x846f826f, 0x857d9972,
+       0x8b7d817d, 0xbf06817d,
+       0xbfa20024, 0xbef600ff,
+       0x01000000, 0xbeee0078,
        0x8078ff78, 0x00000200,
-       0x8078ff78, 0x00000080,
-       0xbef600ff, 0x01000000,
-       0x857d9972, 0x8b7d817d,
-       0xbf06817d, 0xbefd0080,
-       0xbfa2000c, 0xe0500000,
-       0x781d0000, 0xbf890000,
-       0xdac00000, 0x00000000,
-       0x807dff7d, 0x00000080,
-       0x8078ff78, 0x00000080,
-       0xbf0a6f7d, 0xbfa2fff5,
-       0xbfa0000b, 0xe0500000,
-       0x781d0000, 0xbf890000,
-       0xdac00000, 0x00000000,
-       0x807dff7d, 0x00000100,
-       0x8078ff78, 0x00000100,
-       0xbf0a6f7d, 0xbfa2fff5,
-       0xbef80080, 0xbefe00c1,
-       0x857d9972, 0x8b7d817d,
-       0xbf06817d, 0xbfa20002,
-       0xbeff0080, 0xbfa00001,
-       0xbeff00c1, 0xb8ef3b05,
-       0x806f816f, 0x846f826f,
-       0x857d9972, 0x8b7d817d,
-       0xbf06817d, 0xbfa20024,
-       0xbef600ff, 0x01000000,
-       0xbeee0078, 0x8078ff78,
-       0x00000200, 0xbefd0084,
-       0xbf0a6f7d, 0xbfa10050,
+       0xbefd0084, 0xbf0a6f7d,
+       0xbfa10050, 0xe0505000,
+       0x781d0000, 0xe0505080,
+       0x781d0100, 0xe0505100,
+       0x781d0200, 0xe0505180,
+       0x781d0300, 0xbf890000,
+       0x7e008500, 0x7e028501,
+       0x7e048502, 0x7e068503,
+       0x807d847d, 0x8078ff78,
+       0x00000200, 0xbf0a6f7d,
+       0xbfa2ffee, 0xe0505000,
+       0x6e1d0000, 0xe0505080,
+       0x6e1d0100, 0xe0505100,
+       0x6e1d0200, 0xe0505180,
+       0x6e1d0300, 0xbf890000,
+       0xbfa00034, 0xbef600ff,
+       0x01000000, 0xbeee0078,
+       0x8078ff78, 0x00000400,
+       0xbefd0084, 0xbf0a6f7d,
+       0xbfa10012, 0xe0505000,
+       0x781d0000, 0xe0505100,
+       0x781d0100, 0xe0505200,
+       0x781d0200, 0xe0505300,
+       0x781d0300, 0xbf890000,
+       0x7e008500, 0x7e028501,
+       0x7e048502, 0x7e068503,
+       0x807d847d, 0x8078ff78,
+       0x00000400, 0xbf0a6f7d,
+       0xbfa2ffee, 0xb8ef1e06,
+       0x8b6fc16f, 0xbfa1000e,
+       0x846f836f, 0x806f7d6f,
+       0xbefe00c1, 0xbeff0080,
        0xe0505000, 0x781d0000,
-       0xe0505080, 0x781d0100,
-       0xe0505100, 0x781d0200,
-       0xe0505180, 0x781d0300,
        0xbf890000, 0x7e008500,
-       0x7e028501, 0x7e048502,
-       0x7e068503, 0x807d847d,
-       0x8078ff78, 0x00000200,
-       0xbf0a6f7d, 0xbfa2ffee,
+       0x807d817d, 0x8078ff78,
+       0x00000080, 0xbf0a6f7d,
+       0xbfa2fff7, 0xbeff00c1,
        0xe0505000, 0x6e1d0000,
-       0xe0505080, 0x6e1d0100,
-       0xe0505100, 0x6e1d0200,
-       0xe0505180, 0x6e1d0300,
-       0xbf890000, 0xbfa00034,
-       0xbef600ff, 0x01000000,
-       0xbeee0078, 0x8078ff78,
-       0x00000400, 0xbefd0084,
-       0xbf0a6f7d, 0xbfa10012,
-       0xe0505000, 0x781d0000,
-       0xe0505100, 0x781d0100,
-       0xe0505200, 0x781d0200,
-       0xe0505300, 0x781d0300,
-       0xbf890000, 0x7e008500,
-       0x7e028501, 0x7e048502,
-       0x7e068503, 0x807d847d,
-       0x8078ff78, 0x00000400,
-       0xbf0a6f7d, 0xbfa2ffee,
-       0xb8ef1e06, 0x8b6fc16f,
-       0xbfa1000e, 0x846f836f,
-       0x806f7d6f, 0xbefe00c1,
-       0xbeff0080, 0xe0505000,
-       0x781d0000, 0xbf890000,
-       0x7e008500, 0x807d817d,
-       0x8078ff78, 0x00000080,
-       0xbf0a6f7d, 0xbfa2fff7,
-       0xbeff00c1, 0xe0505000,
-       0x6e1d0000, 0xe0505100,
-       0x6e1d0100, 0xe0505200,
-       0x6e1d0200, 0xe0505300,
-       0x6e1d0300, 0xbf890000,
+       0xe0505100, 0x6e1d0100,
+       0xe0505200, 0x6e1d0200,
+       0xe0505300, 0x6e1d0300,
+       0xbf890000, 0xb8f83b05,
+       0x80788178, 0xbf0d9972,
+       0xbfa20002, 0x84788978,
+       0xbfa00001, 0x84788a78,
+       0xb8ee1e06, 0x846e8a6e,
+       0x80786e78, 0x8078ff78,
+       0x00000200, 0x80f8ff78,
+       0x00000050, 0xbef600ff,
+       0x01000000, 0xbefd00ff,
+       0x0000006c, 0x80f89078,
+       0xf428403a, 0xf0000000,
+       0xbf890000, 0x80fd847d,
+       0xbf800000, 0xbe804300,
+       0xbe824302, 0x80f8a078,
+       0xf42c403a, 0xf0000000,
+       0xbf890000, 0x80fd887d,
+       0xbf800000, 0xbe804300,
+       0xbe824302, 0xbe844304,
+       0xbe864306, 0x80f8c078,
+       0xf430403a, 0xf0000000,
+       0xbf890000, 0x80fd907d,
+       0xbf800000, 0xbe804300,
+       0xbe824302, 0xbe844304,
+       0xbe864306, 0xbe884308,
+       0xbe8a430a, 0xbe8c430c,
+       0xbe8e430e, 0xbf06807d,
+       0xbfa1fff0, 0xb980f801,
+       0x00000000, 0xbfbd0000,
        0xb8f83b05, 0x80788178,
        0xbf0d9972, 0xbfa20002,
        0x84788978, 0xbfa00001,
        0x84788a78, 0xb8ee1e06,
        0x846e8a6e, 0x80786e78,
        0x8078ff78, 0x00000200,
-       0x80f8ff78, 0x00000050,
        0xbef600ff, 0x01000000,
-       0xbefd00ff, 0x0000006c,
-       0x80f89078, 0xf428403a,
-       0xf0000000, 0xbf890000,
-       0x80fd847d, 0xbf800000,
-       0xbe804300, 0xbe824302,
-       0x80f8a078, 0xf42c403a,
-       0xf0000000, 0xbf890000,
-       0x80fd887d, 0xbf800000,
-       0xbe804300, 0xbe824302,
-       0xbe844304, 0xbe864306,
-       0x80f8c078, 0xf430403a,
-       0xf0000000, 0xbf890000,
-       0x80fd907d, 0xbf800000,
-       0xbe804300, 0xbe824302,
-       0xbe844304, 0xbe864306,
-       0xbe884308, 0xbe8a430a,
-       0xbe8c430c, 0xbe8e430e,
-       0xbf06807d, 0xbfa1fff0,
-       0xb980f801, 0x00000000,
-       0xbfbd0000, 0xb8f83b05,
-       0x80788178, 0xbf0d9972,
-       0xbfa20002, 0x84788978,
-       0xbfa00001, 0x84788a78,
-       0xb8ee1e06, 0x846e8a6e,
-       0x80786e78, 0x8078ff78,
-       0x00000200, 0xbef600ff,
-       0x01000000, 0xf4205bfa,
+       0xf4205bfa, 0xf0000000,
+       0x80788478, 0xf4205b3a,
        0xf0000000, 0x80788478,
-       0xf4205b3a, 0xf0000000,
-       0x80788478, 0xf4205b7a,
+       0xf4205b7a, 0xf0000000,
+       0x80788478, 0xf4205c3a,
        0xf0000000, 0x80788478,
-       0xf4205c3a, 0xf0000000,
-       0x80788478, 0xf4205c7a,
+       0xf4205c7a, 0xf0000000,
+       0x80788478, 0xf4205eba,
        0xf0000000, 0x80788478,
-       0xf4205eba, 0xf0000000,
-       0x80788478, 0xf4205efa,
+       0xf4205efa, 0xf0000000,
+       0x80788478, 0xf4205e7a,
        0xf0000000, 0x80788478,
-       0xf4205e7a, 0xf0000000,
-       0x80788478, 0xf4205cfa,
+       0xf4205cfa, 0xf0000000,
+       0x80788478, 0xf4205bba,
        0xf0000000, 0x80788478,
+       0xbf890000, 0xb96ef814,
        0xf4205bba, 0xf0000000,
        0x80788478, 0xbf890000,
-       0xb96ef814, 0xf4205bba,
-       0xf0000000, 0x80788478,
-       0xbf890000, 0xb96ef815,
-       0xbefd006f, 0xbefe0070,
-       0xbeff0071, 0xb97bf803,
-       0xb973f801, 0xb8ee3b05,
-       0x806e816e, 0xbf0d9972,
-       0xbfa20002, 0x846e896e,
-       0xbfa00001, 0x846e8a6e,
-       0xb8ef1e06, 0x846f8a6f,
-       0x806e6f6e, 0x806eff6e,
-       0x00000200, 0x806e746e,
-       0x826f8075, 0x8b6fff6f,
-       0x0000ffff, 0xf4085c37,
-       0xf8000050, 0xf4085d37,
-       0xf8000060, 0xf4005e77,
-       0xf8000074, 0xbf890000,
-       0x8b6dff6d, 0x0000ffff,
-       0x8bfe7e7e, 0x8bea6a6a,
-       0xb8eef802, 0xbf0d866e,
-       0xbfa20002, 0xb97af802,
-       0xbe80486c, 0xb97af802,
-       0xbe804a6c, 0xbfb10000,
+       0xb96ef815, 0xbefd006f,
+       0xbefe0070, 0xbeff0071,
+       0xb97bf803, 0xb973f801,
+       0xb8ee3b05, 0x806e816e,
+       0xbf0d9972, 0xbfa20002,
+       0x846e896e, 0xbfa00001,
+       0x846e8a6e, 0xb8ef1e06,
+       0x846f8a6f, 0x806e6f6e,
+       0x806eff6e, 0x00000200,
+       0x806e746e, 0x826f8075,
+       0x8b6fff6f, 0x0000ffff,
+       0xf4085c37, 0xf8000050,
+       0xf4085d37, 0xf8000060,
+       0xf4005e77, 0xf8000074,
+       0xbf890000, 0x8b6dff6d,
+       0x0000ffff, 0x8bfe7e7e,
+       0x8bea6a6a, 0xb8eef802,
+       0xbf0d866e, 0xbfa20002,
+       0xb97af802, 0xbe80486c,
+       0xb97af802, 0xbe804a6c,
+       0xbfb10000, 0xbf9f0000,
        0xbf9f0000, 0xbf9f0000,
        0xbf9f0000, 0xbf9f0000,
-       0xbf9f0000, 0x00000000,
 };
 
 static const uint32_t cwsr_trap_gfx9_4_3_hex[] = {
@@ -3635,7 +3636,7 @@ static const uint32_t cwsr_trap_gfx9_4_3_hex[] = {
 };
 
 static const uint32_t cwsr_trap_gfx12_hex[] = {
-       0xbfa00001, 0xbfa00247,
+       0xbfa00001, 0xbfa0024b,
        0xb0804009, 0xb8f8f804,
        0x9178ff78, 0x00008c00,
        0xb8fbf811, 0x8b6eff78,
@@ -3675,7 +3676,9 @@ static const uint32_t cwsr_trap_gfx12_hex[] = {
        0x8b6dff6d, 0x0000ffff,
        0x8bfe7e7e, 0x8bea6a6a,
        0x85788978, 0xb9783244,
-       0xbe804a6c, 0x8b6dff6d,
+       0xbe804a6c, 0xb8faf802,
+       0xbf0d987a, 0xbfa10001,
+       0xbfb00000, 0x8b6dff6d,
        0x0000ffff, 0xbefa0080,
        0xb97a0151, 0xbeee007e,
        0xbeef007f, 0xbefe0180,
index 18e012e044934feb905189ef9d41eaa7521ab7b9..ac3702b8e3c4e7664dcc65fa87ec7f4327976206 100644 (file)
@@ -97,6 +97,7 @@ var S_STATUS_HALT_MASK                                = SQ_WAVE_STATE_PRIV_HALT_MASK
 var S_SAVE_PC_HI_TRAP_ID_MASK                  = 0xF0000000
 #endif
 
+var SQ_WAVE_STATUS_NO_VGPRS_SHIFT              = 24
 var SQ_WAVE_LDS_ALLOC_LDS_SIZE_SHIFT           = 12
 var SQ_WAVE_LDS_ALLOC_LDS_SIZE_SIZE            = 9
 var SQ_WAVE_GPR_ALLOC_VGPR_SIZE_SIZE           = 8
@@ -451,6 +452,22 @@ L_EXIT_TRAP:
        s_rfe_b64       [ttmp0, ttmp1]
 
 L_SAVE:
+       // If VGPRs have been deallocated then terminate the wavefront.
+       // It has no remaining program to run and cannot save without VGPRs.
+#if ASIC_FAMILY == CHIP_PLUM_BONITO
+       s_bitcmp1_b32   s_save_status, SQ_WAVE_STATUS_NO_VGPRS_SHIFT
+       s_cbranch_scc0  L_HAVE_VGPRS
+       s_endpgm
+L_HAVE_VGPRS:
+#endif
+#if ASIC_FAMILY >= CHIP_GFX12
+       s_getreg_b32    s_save_tmp, hwreg(HW_REG_WAVE_STATUS)
+       s_bitcmp1_b32   s_save_tmp, SQ_WAVE_STATUS_NO_VGPRS_SHIFT
+       s_cbranch_scc0  L_HAVE_VGPRS
+       s_endpgm
+L_HAVE_VGPRS:
+#endif
+
        s_and_b32       s_save_pc_hi, s_save_pc_hi, 0x0000ffff                  //pc[47:32]
        s_mov_b32       s_save_tmp, 0
        s_setreg_b32    hwreg(S_TRAPSTS_HWREG, S_TRAPSTS_SAVE_CONTEXT_SHIFT, 1), s_save_tmp     //clear saveCtx bit