Intel-XE Archive on lore.kernel.org
 help / color / mirror / Atom feed
From: Matthew Brost <matthew.brost@intel.com>
To: Zhanjun Dong <zhanjun.dong@intel.com>
Cc: <intel-xe@lists.freedesktop.org>
Subject: Re: [PATCH v3 2/2] drm/xe/guc: Compress GuC log and CTB dump with zstd
Date: Mon, 29 Jun 2026 21:32:36 -0700	[thread overview]
Message-ID: <akNG5Nv/dp6+m4mP@gsse-cloud1.jf.intel.com> (raw)
In-Reply-To: <20260629172239.2377117-3-zhanjun.dong@intel.com>

On Mon, Jun 29, 2026 at 01:22:39PM -0400, Zhanjun Dong wrote:
> Replace the raw ascii85 dumps of GuC log snapshots with a
> zstd-compressed ascii85 stream and reuse the same helper for CTB
> snapshot output.
> 
> Use zstd streaming with default level-3 parameters and an estimated
> source size of 0 so the compression workspace stays bounded even for
> multi-megabyte logs.
> 
> Add a new Kconfig option DRM_XE_COMPRESS_DUMP to control whether
> GuC log and CTB dumps are compressed with ZSTD. This allows users to
> choose between reduced dump sizes or lower memory usage.
> 

What is the advantage of doing this the KMD vs. getting the dump and
compressing it in user space before sharing it?

> Signed-off-by: Zhanjun Dong <zhanjun.dong@intel.com>
> ---
>  drivers/gpu/drm/xe/Kconfig          |  16 +++
>  drivers/gpu/drm/xe/xe_devcoredump.c |   4 +-
>  drivers/gpu/drm/xe/xe_guc_ct.c      |  19 ++--
>  drivers/gpu/drm/xe/xe_guc_ct.h      |   4 +-
>  drivers/gpu/drm/xe/xe_guc_log.c     | 161 +++++++++++++++++++++++++---
>  drivers/gpu/drm/xe/xe_guc_log.h     |   9 +-
>  6 files changed, 184 insertions(+), 29 deletions(-)
> 
> diff --git a/drivers/gpu/drm/xe/Kconfig b/drivers/gpu/drm/xe/Kconfig
> index 4d7dcaff2b91..3af15a0c14f7 100644
> --- a/drivers/gpu/drm/xe/Kconfig
> +++ b/drivers/gpu/drm/xe/Kconfig
> @@ -33,6 +33,8 @@ config DRM_XE
>  	select ACPI_WMI if X86 && ACPI
>  	select SYNC_FILE
>  	select CRC32
> +	select ZSTD_COMPRESS if DRM_XE_COMPRESS_DUMP
> +	select ZSTD_DECOMPRESS if DRM_XE_COMPRESS_DUMP
>  	select SND_HDA_I915 if SND_HDA_CORE
>  	select CEC_CORE if CEC_NOTIFIER
>  	select VMAP_PFN
> @@ -101,6 +103,20 @@ config DRM_XE_PAGEMAP
>  
>  	  If in doubt say "Y".
>  
> +config DRM_XE_COMPRESS_DUMP
> +	bool "Enable ZSTD compression for error dumps"
> +	depends on DRM_XE
> +	default y

Are tools to decode GuC logs updated to handle this mode?

Matt 

> +	help
> +	  Enable this option to compress GuC log and CTB dumps with ZSTD
> +	  compression. This can significantly reduce the size of error dumps
> +	  but requires additional memory for compression workspace.
> +
> +	  Disable this option if you want to save memory or prefer uncompressed
> +	  error dumps for debugging.
> +
> +	  If in doubt say "Y".
> +
>  config DRM_XE_FORCE_PROBE
>  	string "Force probe xe for selected Intel hardware IDs"
>  	depends on DRM_XE
> diff --git a/drivers/gpu/drm/xe/xe_devcoredump.c b/drivers/gpu/drm/xe/xe_devcoredump.c
> index 5f2b90b18f97..27473c25aad2 100644
> --- a/drivers/gpu/drm/xe/xe_devcoredump.c
> +++ b/drivers/gpu/drm/xe/xe_devcoredump.c
> @@ -115,9 +115,9 @@ static ssize_t __xe_devcoredump_read(char *buffer, ssize_t count,
>  	drm_printf(&p, "\tTile: %d\n", ss->gt->tile->id);
>  
>  	drm_puts(&p, "\n**** GuC Log ****\n");
> -	xe_guc_log_snapshot_print(ss->guc.log, &p);
> +	xe_guc_log_snapshot_print(ss->guc.log, &p, ss->gt);
>  	drm_puts(&p, "\n**** GuC CT ****\n");
> -	xe_guc_ct_snapshot_print(ss->guc.ct, &p);
> +	xe_guc_ct_snapshot_print(ss->guc.ct, &p, ss->gt);
>  
>  	drm_puts(&p, "\n**** Contexts ****\n");
>  	xe_guc_exec_queue_snapshot_print(ss->ge, &p);
> diff --git a/drivers/gpu/drm/xe/xe_guc_ct.c b/drivers/gpu/drm/xe/xe_guc_ct.c
> index 21e0dad9a481..4cedbcdb831e 100644
> --- a/drivers/gpu/drm/xe/xe_guc_ct.c
> +++ b/drivers/gpu/drm/xe/xe_guc_ct.c
> @@ -2094,11 +2094,12 @@ struct xe_guc_ct_snapshot *xe_guc_ct_snapshot_capture(struct xe_guc_ct *ct)
>   * xe_guc_ct_snapshot_print - Print out a given GuC CT snapshot.
>   * @snapshot: GuC CT snapshot object.
>   * @p: drm_printer where it will be printed out.
> + * @gt: GT structure for error reporting
>   *
>   * This function prints out a given GuC CT snapshot object.
>   */
>  void xe_guc_ct_snapshot_print(struct xe_guc_ct_snapshot *snapshot,
> -			      struct drm_printer *p)
> +			      struct drm_printer *p, struct xe_gt *gt)
>  {
>  	if (!snapshot)
>  		return;
> @@ -2112,11 +2113,10 @@ void xe_guc_ct_snapshot_print(struct xe_guc_ct_snapshot *snapshot,
>  		drm_printf(p, "\tg2h outstanding: %d\n",
>  			   snapshot->g2h_outstanding);
>  
> -		if (snapshot->ctb) {
> -			drm_printf(p, "[CTB].length: 0x%zx\n", snapshot->ctb_size);
> -			xe_print_blob_ascii85(p, "[CTB].data", '\n',
> -					      snapshot->ctb, 0, snapshot->ctb_size);
> -		}
> +		if (snapshot->ctb)
> +			xe_guc_print_blob(gt, p, "[CTB]", "[CTB].data",
> +					  &snapshot->ctb, 1,
> +					  snapshot->ctb_size, snapshot->ctb_size);
>  	} else {
>  		drm_puts(p, "CT disabled\n");
>  	}
> @@ -2150,9 +2150,10 @@ void xe_guc_ct_snapshot_free(struct xe_guc_ct_snapshot *snapshot)
>  void xe_guc_ct_print(struct xe_guc_ct *ct, struct drm_printer *p, bool want_ctb)
>  {
>  	struct xe_guc_ct_snapshot *snapshot;
> +	struct xe_gt *gt = ct_to_gt(ct);
>  
>  	snapshot = guc_ct_snapshot_capture(ct, false, want_ctb);
> -	xe_guc_ct_snapshot_print(snapshot, p);
> +	xe_guc_ct_snapshot_print(snapshot, p, gt);
>  	xe_guc_ct_snapshot_free(snapshot);
>  }
>  
> @@ -2250,10 +2251,10 @@ static void ct_dead_print(struct xe_dead_ct *dead)
>  	drm_printf(&lp, "\tTile: %d\n", gt->tile->id);
>  
>  	drm_puts(&lp, "**** GuC Log ****\n");
> -	xe_guc_log_snapshot_print(dead->snapshot_log, &lp);
> +	xe_guc_log_snapshot_print(dead->snapshot_log, &lp, gt);
>  
>  	drm_puts(&lp, "**** GuC CT ****\n");
> -	xe_guc_ct_snapshot_print(dead->snapshot_ct, &lp);
> +	xe_guc_ct_snapshot_print(dead->snapshot_ct, &lp, gt);
>  
>  	drm_puts(&lp, "Done.\n");
>  }
> diff --git a/drivers/gpu/drm/xe/xe_guc_ct.h b/drivers/gpu/drm/xe/xe_guc_ct.h
> index 767365a33dee..2005d1a0cd6c 100644
> --- a/drivers/gpu/drm/xe/xe_guc_ct.h
> +++ b/drivers/gpu/drm/xe/xe_guc_ct.h
> @@ -10,6 +10,7 @@
>  
>  struct drm_printer;
>  struct xe_device;
> +struct xe_gt;
>  
>  int xe_guc_ct_init_noalloc(struct xe_guc_ct *ct);
>  int xe_guc_ct_init(struct xe_guc_ct *ct);
> @@ -24,7 +25,8 @@ void xe_guc_ct_flush_and_stop(struct xe_guc_ct *ct);
>  void xe_guc_ct_fast_path(struct xe_guc_ct *ct);
>  
>  struct xe_guc_ct_snapshot *xe_guc_ct_snapshot_capture(struct xe_guc_ct *ct);
> -void xe_guc_ct_snapshot_print(struct xe_guc_ct_snapshot *snapshot, struct drm_printer *p);
> +void xe_guc_ct_snapshot_print(struct xe_guc_ct_snapshot *snapshot, struct drm_printer *p,
> +			      struct xe_gt *gt);
>  void xe_guc_ct_snapshot_free(struct xe_guc_ct_snapshot *snapshot);
>  void xe_guc_ct_print(struct xe_guc_ct *ct, struct drm_printer *p, bool want_ctb);
>  
> diff --git a/drivers/gpu/drm/xe/xe_guc_log.c b/drivers/gpu/drm/xe/xe_guc_log.c
> index 538d4df0f7aa..d20be1a064f1 100644
> --- a/drivers/gpu/drm/xe/xe_guc_log.c
> +++ b/drivers/gpu/drm/xe/xe_guc_log.c
> @@ -6,6 +6,9 @@
>  #include "xe_guc_log.h"
>  
>  #include <linux/fault-inject.h>
> +#if IS_ENABLED(CONFIG_DRM_XE_COMPRESS_DUMP)
> +#include <linux/zstd.h>
> +#endif
>  
>  #include <linux/utsname.h>
>  #include <drm/drm_managed.h>
> @@ -225,16 +228,148 @@ struct xe_guc_log_snapshot *xe_guc_log_snapshot_capture(struct xe_guc_log *log,
>  	return snapshot;
>  }
>  
> +#if IS_ENABLED(CONFIG_DRM_XE_COMPRESS_DUMP)
> +static void xe_guc_print_blob_compressed(struct xe_gt *gt, struct drm_printer *p,
> +					 const char *length_label, const char *data_label,
> +					 void **chunks, int num_chunks,
> +					 size_t chunk_size, size_t total_size)
> +{
> +	zstd_parameters params;
> +	zstd_cstream *cstream;
> +	zstd_out_buffer outbuf;
> +	size_t wksp_size, dst_size;
> +	size_t remain, ret;
> +	void *wksp = NULL;
> +	char *dst = NULL;
> +	int i;
> +
> +	/*
> +	 * Pass 0 as estimated_src_size to use zstd level-3 defaults
> +	 * (windowLog=18, 256KB window). Passing the full total_size would
> +	 * cause zstd to inflate windowLog proportionally (e.g. windowLog=24
> +	 * for a 19MB log), making the workspace tens of MB.  With 0, the
> +	 * workspace stays ~1.5MB, well within kmalloc limits.
> +	 */
> +	params = zstd_get_params(3, 0);
> +
> +	wksp_size = zstd_cstream_workspace_bound(&params.cParams);
> +	wksp = kvzalloc(wksp_size, GFP_KERNEL);
> +	if (!wksp)
> +		goto err;
> +
> +	dst_size = zstd_compress_bound(total_size);
> +	dst = kvzalloc(dst_size, GFP_KERNEL);
> +	if (!dst)
> +		goto err;
> +
> +	cstream = zstd_init_cstream(&params, 0, wksp, wksp_size);
> +	if (!cstream) {
> +		xe_gt_err(gt, "failed to init zstd cstream\n");
> +		goto err;
> +	}
> +
> +	outbuf.dst = dst;
> +	outbuf.size = dst_size;
> +	outbuf.pos = 0;
> +
> +	remain = total_size;
> +	for (i = 0; i < num_chunks && remain; i++) {
> +		size_t size = min(chunk_size, remain);
> +		zstd_in_buffer inbuf = { .src = chunks[i], .size = size, .pos = 0 };
> +
> +		while (inbuf.pos < inbuf.size) {
> +			ret = zstd_compress_stream(cstream, &outbuf, &inbuf);
> +			if (zstd_is_error(ret)) {
> +				xe_gt_err(gt, "zstd_compress_stream failed: %s\n",
> +					  zstd_get_error_name(ret));
> +				goto err;
> +			}
> +		}
> +		remain -= size;
> +	}
> +
> +	do {
> +		ret = zstd_end_stream(cstream, &outbuf);
> +		if (zstd_is_error(ret)) {
> +			xe_gt_err(gt, "zstd_end_stream failed: %s\n",
> +				  zstd_get_error_name(ret));
> +			goto err;
> +		}
> +	} while (ret > 0 && outbuf.pos < outbuf.size);
> +
> +	drm_printf(p, "%s.length: 0x%zx -> 0x%zx Algo: ZSTD\n",
> +		   length_label, total_size, outbuf.pos);
> +	xe_print_blob_ascii85(p, data_label, '\n', dst, 0,
> +			      DIV_ROUND_UP(outbuf.pos, sizeof(u32)) * sizeof(u32));
> +	goto out;
> +
> +err:
> +	drm_printf(p, "%s.length: 0x%zx Algo: ERROR (compression failed)\n",
> +		   length_label, total_size);
> +out:
> +	kvfree(dst);
> +	kvfree(wksp);
> +}
> +#endif
> +
>  /**
> - * xe_guc_log_snapshot_print - dump a previously saved copy of the GuC log to some useful location
> - * @snapshot: a snapshot of the GuC log
> - * @p: the printer object to output to
> + * xe_guc_print_blob - Print binary data (compressed or uncompressed) as ascii85
> + * @gt: GT structure for error reporting
> + * @p: drm_printer for output
> + * @length_label: label prefix for the length/metadata line (e.g. "[LOG]" or "[CTB]")
> + * @data_label: label for the ascii85 blob (e.g. "[LOG].data" or "[CTB].data")
> + * @chunks: array of pointers to source data chunks
> + * @num_chunks: number of elements in @chunks
> + * @chunk_size: nominal size of each chunk; the last chunk may hold fewer valid bytes
> + * @total_size: total uncompressed byte count across all chunks
> + *
> + * If CONFIG_DRM_XE_COMPRESS_DUMP is enabled, compresses @total_size bytes using
> + * zstd level-3 streaming compression and prints the result via @p using ascii85
> + * encoding. Otherwise, directly prints the uncompressed data.
> + * On failure the function returns without printing the data blob.
>   */
> -void xe_guc_log_snapshot_print(struct xe_guc_log_snapshot *snapshot, struct drm_printer *p)
> +void xe_guc_print_blob(struct xe_gt *gt __maybe_unused, struct drm_printer *p,
> +		       const char *length_label, const char *data_label,
> +		       void **chunks, int num_chunks,
> +		       size_t chunk_size, size_t total_size)
>  {
> +#if IS_ENABLED(CONFIG_DRM_XE_COMPRESS_DUMP)
> +	xe_guc_print_blob_compressed(gt, p, length_label, data_label,
> +				     chunks, num_chunks, chunk_size, total_size);
> +#else
> +	/* Compression disabled, output uncompressed data directly */
>  	size_t remain;
>  	int i;
>  
> +	drm_printf(p, "%s.length: 0x%zx Algo: RAW\n", length_label, total_size);
> +
> +	/*
> +	 * Print each chunk directly without merging into a single buffer.
> +	 * xe_print_blob_ascii85() supports multiple calls - use suffix=0
> +	 * for all but the last chunk to omit the newline between chunks.
> +	 */
> +	remain = total_size;
> +	for (i = 0; i < num_chunks && remain; i++) {
> +		size_t size = min(chunk_size, remain);
> +		const char *prefix = i ? NULL : data_label;
> +		/* Add suffix only on the last chunk */
> +		char suffix = (remain == size) ? '\n' : 0;
> +
> +		xe_print_blob_ascii85(p, prefix, suffix, chunks[i], 0,
> +				      DIV_ROUND_UP(size, sizeof(u32)) * sizeof(u32));
> +		remain -= size;
> +	}
> +#endif
> +}
> +
> +/**
> + * xe_guc_log_snapshot_print - dump a previously saved copy of the GuC log to some useful location
> + * @snapshot: a snapshot of the GuC log
> + * @p: the printer object to output to
> + */
> +void xe_guc_log_snapshot_print(struct xe_guc_log_snapshot *snapshot, struct drm_printer *p,
> +			       struct xe_gt *gt)
> +{
>  	if (!snapshot) {
>  		drm_printf(p, "GuC log snapshot not allocated!\n");
>  		return;
> @@ -248,16 +383,9 @@ void xe_guc_log_snapshot_print(struct xe_guc_log_snapshot *snapshot, struct drm_
>  	drm_printf(p, "GuC timestamp: 0x%08llX [%llu]\n", snapshot->stamp, snapshot->stamp);
>  	drm_printf(p, "Log level: %u\n", snapshot->level);
>  
> -	drm_printf(p, "[LOG].length: 0x%zx\n", snapshot->size);
> -	remain = snapshot->size;
> -	for (i = 0; i < snapshot->num_chunks; i++) {
> -		size_t size = min(GUC_LOG_CHUNK_SIZE, remain);
> -		const char *prefix = i ? NULL : "[LOG].data";
> -		char suffix = i == snapshot->num_chunks - 1 ? '\n' : 0;
> -
> -		xe_print_blob_ascii85(p, prefix, suffix, snapshot->copy[i], 0, size);
> -		remain -= size;
> -	}
> +	xe_guc_print_blob(gt, p, "[LOG]", "[LOG].data",
> +			  snapshot->copy, snapshot->num_chunks,
> +			  GUC_LOG_CHUNK_SIZE, snapshot->size);
>  }
>  
>  static inline void lfd_output_binary(struct drm_printer *p, char *buf, int buf_size)
> @@ -598,12 +726,13 @@ void xe_guc_log_print_dmesg(struct xe_guc_log *log)
>  void xe_guc_log_print(struct xe_guc_log *log, struct drm_printer *p)
>  {
>  	struct xe_guc_log_snapshot *snapshot;
> +	struct xe_gt *gt = log_to_gt(log);
>  
>  	drm_printf(p, "**** GuC Log ****\n");
>  
>  	snapshot = xe_guc_log_snapshot_capture(log, false);
> -	drm_printf(p, "CS reference clock: %u\n", log_to_gt(log)->info.reference_clock);
> -	xe_guc_log_snapshot_print(snapshot, p);
> +	drm_printf(p, "CS reference clock: %u\n", gt->info.reference_clock);
> +	xe_guc_log_snapshot_print(snapshot, p, gt);
>  	xe_guc_log_snapshot_free(snapshot);
>  }
>  
> diff --git a/drivers/gpu/drm/xe/xe_guc_log.h b/drivers/gpu/drm/xe/xe_guc_log.h
> index 4649a260755e..3e97647ee8b5 100644
> --- a/drivers/gpu/drm/xe/xe_guc_log.h
> +++ b/drivers/gpu/drm/xe/xe_guc_log.h
> @@ -11,6 +11,7 @@
>  
>  struct drm_printer;
>  struct xe_device;
> +struct xe_gt;
>  
>  #if IS_ENABLED(CONFIG_DRM_XE_DEBUG_GUC)
>  #define XE_GUC_LOG_EVENT_DATA_BUFFER_SIZE	SZ_16M
> @@ -59,7 +60,13 @@ void xe_guc_log_print(struct xe_guc_log *log, struct drm_printer *p);
>  void xe_guc_log_print_lfd(struct xe_guc_log *log, struct drm_printer *p);
>  void xe_guc_log_print_dmesg(struct xe_guc_log *log);
>  struct xe_guc_log_snapshot *xe_guc_log_snapshot_capture(struct xe_guc_log *log, bool atomic);
> -void xe_guc_log_snapshot_print(struct xe_guc_log_snapshot *snapshot, struct drm_printer *p);
> +
> +void xe_guc_print_blob(struct xe_gt *gt, struct drm_printer *p, const char *length_label,
> +		       const char *data_label, void **chunks, int num_chunks,
> +		       size_t chunk_size, size_t total_size);
> +
> +void xe_guc_log_snapshot_print(struct xe_guc_log_snapshot *snapshot, struct drm_printer *p,
> +			       struct xe_gt *gt);
>  void xe_guc_log_snapshot_free(struct xe_guc_log_snapshot *snapshot);
>  
>  static inline u32
> -- 
> 2.34.1
> 

  reply	other threads:[~2026-06-30  4:32 UTC|newest]

Thread overview: 11+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2026-06-29 17:22 [PATCH v3 0/2] drm/xe/guc: shrink and compress GuC snapshot dumps Zhanjun Dong
2026-06-29 17:22 ` [PATCH v3 1/2] drm/xe: Remove extra newlines from LRC snapshot dump Zhanjun Dong
2026-06-29 17:22 ` [PATCH v3 2/2] drm/xe/guc: Compress GuC log and CTB dump with zstd Zhanjun Dong
2026-06-30  4:32   ` Matthew Brost [this message]
2026-06-30 17:35     ` Dong, Zhanjun
2026-07-13 18:13       ` Dong, Zhanjun
2026-07-13 18:29       ` Matthew Brost
2026-06-30  1:26 ` ✗ CI.checkpatch: warning for drm/xe/guc: shrink and compress GuC snapshot dumps Patchwork
2026-06-30  1:27 ` ✓ CI.KUnit: success " Patchwork
2026-06-30  2:02 ` ✓ Xe.CI.BAT: " Patchwork
2026-06-30 14:34 ` ✗ Xe.CI.FULL: failure " Patchwork

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=akNG5Nv/dp6+m4mP@gsse-cloud1.jf.intel.com \
    --to=matthew.brost@intel.com \
    --cc=intel-xe@lists.freedesktop.org \
    --cc=zhanjun.dong@intel.com \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox