Skip to contentSkip to conversation

igt/gem_workarounds: Test all types of workarounds

PATCH Superseded Oscar Mateo 3 participants 7 messages Oct 11, 2017 – Oct 30, 2017

v2 of 5 Superseded Latest is v5 →

lore
Share & tools
lore
#0 lore raw

PATCH v2 1/1 igt/gem_workarounds: Test all types of workarounds

Apart from context based workarounds, we can now also test for global MMIO and whitelisting ones.

Do take into account that this test does not guarantee that all known WAs for a given platform are applied. It only checks that the WAs the kernel does know about are correctly applied (e.g. they didn't get lost on a GPU reset or a suspend/resume).

v2:
- Do not wait for the GPU unnecessarily (Chris)
- Make a comment that this tests only looks for regressions (Chris)

Cc Chris Wilson <[email protected]>
Cc Mika Kuoppala <[email protected]>
---
 tests/gem_workarounds.c | 185 ++++++++++++++++++++++++++++++++++++++----------
 1 file changed, 147 insertions(+), 38 deletions(-)
tests/gem_workarounds.c +147 -38
index 7b99961..69174d9 100644
@@ -28,6 +28,7 @@
 #include "igt.h"
 
 #include <fcntl.h>
+#include <ctype.h>
 
 #define PAGE_SIZE 4096
 #define PAGE_ALIGN(x) ALIGN(x, PAGE_SIZE)
@@ -62,8 +63,14 @@ static struct write_only_list {
 	 */
 };
 
-static struct intel_wa_reg *wa_regs;
-static int num_wa_regs;
+static struct intel_wa_reg *ctx_wa_regs;
+static int num_ctx_wa_regs;
+
+static struct intel_wa_reg *mmio_wa_regs;
+static int num_mmio_wa_regs;
+
+static struct intel_wa_reg *whitelist_wa_regs;
+static int num_whitelist_wa_regs;
 
 static bool write_only(const uint32_t addr)
 {
@@ -82,7 +89,7 @@ static bool write_only(const uint32_t addr)
 
 #define MI_STORE_REGISTER_MEM (0x24 << 23)
 
-static int workaround_fail_count(int fd, uint32_t ctx)
+static int ctx_workarounds_fail_count(int fd, uint32_t ctx)
 {
 	struct drm_i915_gem_exec_object2 obj[2];
 	struct drm_i915_gem_relocation_entry *reloc;
@@ -91,13 +98,16 @@ static int workaround_fail_count(int fd, uint32_t ctx)
 	uint32_t *base, *out;
 	int fail_count = 0;
 
-	reloc = calloc(num_wa_regs, sizeof(*reloc));
+	if (!num_ctx_wa_regs)
+		return 0;
+
+	reloc = calloc(num_ctx_wa_regs, sizeof(*reloc));
 	igt_assert(reloc);
 
-	result_sz = 4 * num_wa_regs;
+	result_sz = 4 * num_ctx_wa_regs;
 	result_sz = PAGE_ALIGN(result_sz);
 
-	batch_sz = 16 * num_wa_regs + 4;
+	batch_sz = 16 * num_ctx_wa_regs + 4;
 	batch_sz = PAGE_ALIGN(batch_sz);
 
 	memset(obj, 0, sizeof(obj));
@@ -105,12 +115,12 @@ static int workaround_fail_count(int fd, uint32_t ctx)
 	gem_set_caching(fd, obj[0].handle, I915_CACHING_CACHED);
 	obj[1].handle = gem_create(fd, batch_sz);
 	obj[1].relocs_ptr = to_user_pointer(reloc);
-	obj[1].relocation_count = num_wa_regs;
+	obj[1].relocation_count = num_ctx_wa_regs;
 
 	out = base = gem_mmap__cpu(fd, obj[1].handle, 0, batch_sz, PROT_WRITE);
-	for (int i = 0; i < num_wa_regs; i++) {
+	for (int i = 0; i < num_ctx_wa_regs; i++) {
 		*out++ = MI_STORE_REGISTER_MEM | ((gen >= 8 ? 4 : 2) - 2);
-		*out++ = wa_regs[i].addr;
+		*out++ = ctx_wa_regs[i].addr;
 		reloc[i].target_handle = obj[0].handle;
 		reloc[i].offset = (out - base) * sizeof(*out);
 		reloc[i].delta = i * sizeof(uint32_t);
@@ -134,20 +144,20 @@ static int workaround_fail_count(int fd, uint32_t ctx)
 	igt_debug("Address\tval\t\tmask\t\tread\t\tresult\n");
 
 	out = gem_mmap__cpu(fd, obj[0].handle, 0, result_sz, PROT_READ);
-	for (int i = 0; i < num_wa_regs; i++) {
+	for (int i = 0; i < num_ctx_wa_regs; i++) {
 		const bool ok =
-			(wa_regs[i].value & wa_regs[i].mask) ==
-			(out[i] & wa_regs[i].mask);
+			(ctx_wa_regs[i].value & ctx_wa_regs[i].mask) ==
+			(out[i] & ctx_wa_regs[i].mask);
 		char buf[80];
 
 		snprintf(buf, sizeof(buf),
 			 "0x%05X\t0x%08X\t0x%08X\t0x%08X",
-			 wa_regs[i].addr, wa_regs[i].value, wa_regs[i].mask,
+			 ctx_wa_regs[i].addr, ctx_wa_regs[i].value, ctx_wa_regs[i].mask,
 			 out[i]);
 
 		if (ok) {
 			igt_debug("%s\tOK\n", buf);
-		} else if (write_only(wa_regs[i].addr)) {
+		} else if (write_only(ctx_wa_regs[i].addr)) {
 			igt_debug("%s\tIGNORED (w/o)\n", buf);
 		} else {
 			igt_warn("%s\tFAIL\n", buf);
@@ -163,6 +173,49 @@ static int workaround_fail_count(int fd, uint32_t ctx)
 	return fail_count;
 }
 
+static int mmio_workarounds_fail_count(struct intel_wa_reg *wa_regs, int num_wa_regs)
+{
+	int i, fail_count = 0;
+
+	if (!num_wa_regs)
+		return 0;
+
+	igt_debug("Address\tval\t\tmask\t\tread\t\tresult\n");
+
+	for (i = 0; i < num_wa_regs; ++i) {
+		const uint32_t val = intel_register_read(wa_regs[i].addr);
+		const bool ok = (wa_regs[i].value & wa_regs[i].mask) ==
+			(val & wa_regs[i].mask);
+
+		igt_debug("0x%05X\t0x%08X\t0x%08X\t0x%08X\t%s\n",
+			  wa_regs[i].addr, wa_regs[i].value,
+			  wa_regs[i].mask, val, ok ? "OK" : "FAIL");
+
+		if (write_only(wa_regs[i].addr))
+			continue;
+
+		if (!ok) {
+			igt_warn("0x%05X\t0x%08X\t0x%08X\t0x%08X\t%s\n",
+				 wa_regs[i].addr, wa_regs[i].value,
+				 wa_regs[i].mask, val, ok ? "OK" : "FAIL");
+			fail_count++;
+		}
+	}
+
+	return fail_count;
+}
+
+static int workarounds_fail_count(int fd, uint32_t ctx)
+{
+	int fail_count = 0;
+
+	fail_count += ctx_workarounds_fail_count(fd, ctx);
+	fail_count += mmio_workarounds_fail_count(mmio_wa_regs, num_mmio_wa_regs);
+	fail_count += mmio_workarounds_fail_count(whitelist_wa_regs, num_whitelist_wa_regs);
+
+	return fail_count;
+}
+
 static int reopen(int fd)
 {
 	char path[256];
@@ -185,7 +238,7 @@ static void check_workarounds(int fd, enum operation op, unsigned int flags)
 	if (flags & CONTEXT)
 		ctx = gem_context_create(fd);
 
-	igt_assert_eq(workaround_fail_count(fd, ctx), 0);
+	igt_assert_eq(workarounds_fail_count(fd, ctx), 0);
 
 	switch (op) {
 	case GPU_RESET:
@@ -209,7 +262,7 @@ static void check_workarounds(int fd, enum operation op, unsigned int flags)
 		igt_assert(0);
 	}
 
-	igt_assert_eq(workaround_fail_count(fd, ctx), 0);
+	igt_assert_eq(workarounds_fail_count(fd, ctx), 0);
 
 	if (flags & CONTEXT)
 		gem_context_destroy(fd, ctx);
@@ -217,6 +270,62 @@ static void check_workarounds(int fd, enum operation op, unsigned int flags)
 		close(fd);
 }
 
+static bool is_empty(const char *s)
+{
+	while (*s != '\0') {
+		if (!isspace(*s))
+			return false;
+		s++;
+	}
+
+	return true;
+}
+
+static char *skip_preamble(char *s, const char *preamble)
+{
+	while (*s == *preamble) {
+		s++;
+		preamble++;
+	}
+
+	return s;
+}
+
+static void read_workarounds(FILE *file, const char *preamble, struct intel_wa_reg **regs, int *num)
+{
+	char *header, *line = NULL;
+	size_t line_size;
+	struct intel_wa_reg *wa_regs;
+	int num_wa_regs;
+	int i = 0;
+
+	igt_assert(getline(&line, &line_size, file) > 0);
+	igt_debug("i915_wa_registers: %s", line);
+	header = skip_preamble(line, preamble);
+	sscanf(header, " workarounds applied: %d", &num_wa_regs);
+
+	wa_regs = malloc(num_wa_regs * sizeof(*wa_regs));
+
+	while (getline(&line, &line_size, file) > 0) {
+		if (is_empty(line))
+			break;
+
+		igt_debug("%s", line);
+		if (sscanf(line, "0x%X: 0x%08X, mask: 0x%08X",
+			   &wa_regs[i].addr,
+			   &wa_regs[i].value,
+			   &wa_regs[i].mask) == 3)
+			i++;
+	}
+
+	igt_assert_lte(i, num_wa_regs);
+
+	*regs = wa_regs;
+	*num = num_wa_regs;
+
+	free(line);
+}
+
 igt_main
 {
 	int device = -1;
@@ -241,39 +350,34 @@ igt_main
 	}, *m;
 
 	igt_fixture {
+		struct pci_device *pci_dev;
 		FILE *file;
-		char *line = NULL;
-		size_t line_size;
-		int i, fd;
+		int fd;
 
 		device = drm_open_driver(DRIVER_INTEL);
 		igt_require_gem(device);
 
 		gen = intel_gen(intel_get_drm_devid(device));
 
+		pci_dev = intel_get_pci_device();
+		igt_require(pci_dev);
+
+		intel_register_access_init(pci_dev, 0, device);
+
 		fd = igt_debugfs_open(device, "i915_wa_registers", O_RDONLY);
 		file = fdopen(fd, "r");
-		igt_assert(getline(&line, &line_size, file) > 0);
-		igt_debug("i915_wa_registers: %s", line);
-		sscanf(line, "Workarounds applied: %d", &num_wa_regs);
-		igt_require(num_wa_regs > 0);
-
-		wa_regs = malloc(num_wa_regs * sizeof(*wa_regs));
-		igt_assert(wa_regs);
-
-		i = 0;
-		while (getline(&line, &line_size, file) > 0) {
-			igt_debug("%s", line);
-			if (sscanf(line, "0x%X: 0x%08X, mask: 0x%08X",
-				   &wa_regs[i].addr,
-				   &wa_regs[i].value,
-				   &wa_regs[i].mask) == 3)
-				i++;
-		}
 
-		igt_assert_lte(i, num_wa_regs);
+		/*
+		 * This test relies on the list of workarounds the kernel says
+		 * have been applied and it only checks that those are (indeed)
+		 * correctly applied. It does not report whether the system has
+		 * applied all known workarounds for a fiven platform.
+		 */
+		read_workarounds(file, "Context", &ctx_wa_regs, &num_ctx_wa_regs);
+		read_workarounds(file, "MMIO", &mmio_wa_regs, &num_mmio_wa_regs);
+		read_workarounds(file, "Whitelist", &whitelist_wa_regs, &num_whitelist_wa_regs);
+		igt_require(num_ctx_wa_regs + num_mmio_wa_regs + num_whitelist_wa_regs > 0);
 
-		free(line);
 		fclose(file);
 		close(fd);
 	}
@@ -284,4 +388,9 @@ igt_main
 				check_workarounds(device, op->op, m->flags);
 		}
 	}
+
+	igt_fixture {
+		free(ctx_wa_regs);
+		free(mmio_wa_regs);
+	}
 }
Signature · 6 lines
1.9.1

_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
#1 lore raw

Quoting Oscar Mateo (2017-10-11 19:15:40)

Quoted 10 lines
tests/gem_workarounds.c +2 -3
@@ -241,39 +350,34 @@ igt_main
        }, *m;
 
        igt_fixture {
+               struct pci_device *pci_dev;
                FILE *file;
-               char *line = NULL;
-               size_t line_size;
-               int i, fd;
+               int fd;

tests/gem_workarounds.c igt_main · line 365

Quoted 10 lines
tests/gem_workarounds.c +4 -0
 
                device = drm_open_driver(DRIVER_INTEL);
                igt_require_gem(device);
 
                gen = intel_gen(intel_get_drm_devid(device));
 
+               pci_dev = intel_get_pci_device();
+               igt_require(pci_dev);
+
+               intel_register_access_init(pci_dev, 0, device);

intel_register_access_init() takes i915-user-forcewake. Can we limit the register access to just the mmio tests?
-Chris
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx

#2 lore raw
In reply to Oscar Mateo

Quoting Oscar Mateo (2017-10-11 19:15:40)

Quoted 7 lines

Apart from context based workarounds, we can now also test for global MMIO and whitelisting ones.

Do take into account that this test does not guarantee that all known WAs for a given platform are applied. It only checks that the WAs the kernel does know about are correctly applied (e.g. they didn't get lost on a GPU reset or a suspend/resume).

Can we pass in a wa_regs.txt (manual control) instead of using i915_wa_regs (regression testing)?

I want to record the wa_regs from the end of the series and confirm that the registers have the same values at the beginning of the series. We can't check if regs have disappeared in between, but that way we can do a quick check they haven't changed value.
-Chris
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx

#3 lore raw
In reply to Oscar Mateo
Quoted 8 lines · Oscar Mateo

Apart from context based workarounds, we can now also test for global MMIO and whitelisting ones.

Do take into account that this test does not guarantee that all known WAs for a given platform are applied. It only checks that the WAs the kernel does know about are correctly applied (e.g. they didn't get lost on a GPU reset or a suspend/resume).

Quoted 7 lines

v2:
- Do not wait for the GPU unnecessarily (Chris)
- Make a comment that this tests only looks for regressions (Chris)

Cc Chris Wilson <[email protected]>
Cc Mika Kuoppala <[email protected]>

Subjectprefix didn't contain i-g-t so CI did not pick this up.

Signature · 5 lines
Petri Latvala
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
#4 lore raw
Quoted 13 lines · Petri Latvala
Quoted 12 lines · Oscar Mateo

Apart from context based workarounds, we can now also test for global MMIO and whitelisting ones.

Do take into account that this test does not guarantee that all known WAs for a given platform are applied. It only checks that the WAs the kernel does know about are correctly applied (e.g. they didn't get lost on a GPU reset or a suspend/resume).

v2:
- Do not wait for the GPU unnecessarily (Chris)
- Make a comment that this tests only looks for regressions (Chris)

Quoted 6 lines
Quoted 3 lines
Cc Chris Wilson <[email protected]>
Cc Mika Kuoppala <[email protected]>

Subjectprefix didn't contain i-g-t so CI did not pick this up.

Thanks for the heads up, Petri. I forgot to add the label, but it doesn't matter that much because this patch is going to fail CI until the companion patches make it to the KMD (is there a BKM to deal with this situation?).

Thanks,
Oscar
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx

#5 lore raw
Quoted 20 lines · Oscar Mateo
Quoted 17 lines · Petri Latvala
Quoted 15 lines · Oscar Mateo

Apart from context based workarounds, we can now also test for global MMIO and whitelisting ones.

Do take into account that this test does not guarantee that all known WAs for a given platform are applied. It only checks that the WAs the kernel does know about are correctly applied (e.g. they didn't get lost on a GPU reset or a suspend/resume).

v2:
- Do not wait for the GPU unnecessarily (Chris)
- Make a comment that this tests only looks for regressions (Chris)

Cc Chris Wilson <[email protected]>
Cc Mika Kuoppala <[email protected]>
Quoted 6 lines
Quoted 2 lines

Subjectprefix didn't contain i-g-t so CI did not pick this up.

Thanks for the heads up, Petri. I forgot to add the label, but it doesn't matter that much because this patch is going to fail CI until the companion patches make it to the KMD (is there a BKM to deal with this situation?).

That still tests if it builds in the CI environment, whether it affects other tests, and shows how the failure behaves.

How is the test supposed to behave on older kernels, like stable releases (4.4.x, 4.12.x, etc)? Is "fail" proper for kernels without the KMD changes required, instead of producing a skip? If that is the case, the best story is to keep sending the patch towards CI, collect reviews, possibly resend for one more CI run after kernel changes have landed to see lights turn green, land tests. If the test is behaving well with "broken" kernels it can even be landed before the kernel changes, if the used interfaces already exist in kernel upstream.

Signature · 5 lines
Petri Latvala
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
#6 lore raw
Quoted 27 lines · Petri Latvala
Quoted 15 lines · Oscar Mateo
Quoted 13 lines · Petri Latvala
Quoted 12 lines · Oscar Mateo

Apart from context based workarounds, we can now also test for global MMIO and whitelisting ones.

Do take into account that this test does not guarantee that all known WAs for a given platform are applied. It only checks that the WAs the kernel does know about are correctly applied (e.g. they didn't get lost on a GPU reset or a suspend/resume).

v2:
- Do not wait for the GPU unnecessarily (Chris)
- Make a comment that this tests only looks for regressions (Chris)

Quoted 8 lines
Quoted 5 lines
Quoted 3 lines
Cc Chris Wilson <[email protected]>
Cc Mika Kuoppala <[email protected]>

Subjectprefix didn't contain i-g-t so CI did not pick this up.

Thanks for the heads up, Petri. I forgot to add the label, but it doesn't matter that much because this patch is going to fail CI until the companion patches make it to the KMD (is there a BKM to deal with this situation?).

That still tests if it builds in the CI environment, whether it affects other tests, and shows how the failure behaves.

Quoted 8 lines

How is the test supposed to behave on older kernels, like stable releases (4.4.x, 4.12.x, etc)? Is "fail" proper for kernels without the KMD changes required, instead of producing a skip? If that is the case, the best story is to keep sending the patch towards CI, collect reviews, possibly resend for one more CI run after kernel changes have landed to see lights turn green, land tests. If the test is behaving well with "broken" kernels it can even be landed before the kernel changes, if the used interfaces already exist in kernel upstream.

Got it. Thank you, Petri. I can modify the test so that it keeps working with older kernels, which looks like the best possible approach.
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx