Re: [PATCH 4/4] drm/xe/guc: Expose raw access to GuC log over debugfs

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On 5/14/2024 13:41, Michal Wajdeczko wrote:
On 14.05.2024 20:13, John Harrison wrote:
On 5/14/2024 07:58, Michal Wajdeczko wrote:
On 13.05.2024 18:53, John Harrison wrote:
On 5/12/2024 08:36, Michal Wajdeczko wrote:
We already provide the content of the GuC log in debugsfs, but it
is in a text format where each log dword is printed as hexadecimal
number, which does not scale well with large GuC log buffers.

To allow more efficient access to the GuC log, which could benefit
our CI systems, expose raw binary log data.  In addition to less
overhead in preparing text based GuC log file, the new GuC log file
in binary format is also almost 3x smaller.

Any existing script that expects the GuC log buffer in text format
can use command like below to convert from new binary format:

      hexdump -e '4/4 "0x%08x " "\n"'

but this shouldn't be the case as most decoders expect GuC log data
in binary format.
I strongly disagree with this.

Efficiency and file size is not an issue when accessing the GuC log via
debugfs on actual hardware.
to some extend it is as CI team used to refuse to collect GuC logs after
each executed test just because of it's size
I've never heard that argument. I've heard many different arguments but
not one about file size. The default GuC log size is pretty tiny. So
size really is not an issue.
so it's tiny or 16MB as you mention below ?
The default size is tiny. The maximum allowed is 16MB. By default, you get tiny logs. When a developer is debugging a specific issue and needs larger logs, they can bump the size up to 16MB.


It is an issue when dumping via dmesg but
you definitely should not be dumping binary data to dmesg. Whereas,
not following here - this is debugfs specific, not a dmesg printer
Except that it is preferable to have common code for both if at all
possible.
but here, for debugfs, it's almost no code, it's 1-liner thanks to using
generic helpers, so there is really nothing to share as common code

note that with this separate raw access to guc log over debugfs, you can
further customize xe_guc_log_dump() function for dmesg output [2]
without worrying about impact in generating output to debugfs

[2] https://patchwork.freedesktop.org/series/133349/
Or, we could put all this extra effort into doing something with tangible benefit. I've probably already spent more time arguing about this patch than it took to implement it. Time I would much rather be doing something useful with.

And my point was that the dump size is only relevant for dmesg. For debugfs, the size simply does not matter. So trying to optimise the debugfs dump size but with a downside of making it more difficult to use and more susceptible to issues is a bad trade off that we should not be making.


dumping in binary data is much more dangerous and liable to corruption
because some tool along the way tries to convert to ASCII, or truncates
at the first zero, etc. We request GuC logs be sent by end users,
customer bug reports, etc. all doing things that we have no control
over.
hmm, how "cp gt0/uc/guc_log_raw FILE" could end with a corrupted file ?
Because someone then tries to email it, or attach it or copy it via
Windows or any number of other ways in which a file can get munged.
no comment

Converting the hexdump back to binary is trivial for those tools which
require it. If you follow the acquisition and decoding instructions on
the wiki page then it is all done for you automatically.
I'm afraid I don't know where this wiki page is, but I do know that hex
conversion dance is not needed for me to get decoded GuC log the way I
used to do
Look for the 'GuC Debug Logs' page on the developer wiki. It's pretty
easy to find.
ok, found it

btw, it says "Actual log size will be significantly more (about 50MB) as
there are multiple sections."
16MB debug log, 1MB crash dump, 1MB register capture -> 18MB actual data size, expands to about 50MB as an ASCII hexdump.


These patches are trying to solve a problem which does not exist and are
going to make working with GuC logs harder and more error prone.
it at least solves the problem of currently super inefficient way of
generating the GuC log in text format.

it also opens other opportunities to develop tools that could monitor or
capture GuC log independently on  top of what driver is able to offer
today (on i915 there was guc-log-relay, but it was broken for long time,
not sure what are the plans for Xe)

also still not sure how it can be more error prone.
As already explained, the plan is move to LFD - an extensible,
streamable, logging format. Any non-trivial effort that is not helping
to move to LFD is not worth the effort.
which part from my series was non-trivial ?
The doing it properly part.

If you want a functional streaming interface then you will need a lot more than a backdoor access to the GuC log memory buffer. You will need all the user land code to interpret it, do the streaming, cope with wrap-arounds, etc. Effort which would be more usefully put towards implementing LFDs because that gives you all of that and much, much, more.


On the other hand, there are many other issues with GuC logs that it
would be useful to solves - including extra meta data, reliable output
via dmesg, continuous streaming, pre-sizing the debugfs file to not have
to generate it ~12 times for a single read, etc.
this series actually solves last issue but in a bit different way (we
even don't need to generate full GuC log dump at all if we would like to
capture only part of the log if we know where to look)
No, it doesn't solve it. Your comment below suggests it will be read in
4KB chunks.
chunks will be 4K if we stick to proposed here simple_read_from_iomem()
that initially uses hardcoded 4K chunks, but we could either modify it
to use larger chunks by default or extend it to take additional param,
or promote more powerful copy_to_user_fromio() from SOUND
Which is yet more effort to still solve the problem in the wrong manner.

If you are trying to implement streaming logs then we should do that via LFDs as that is a much simpler debugfs interface and only requires 'cat' on the userland side. If you are trying to solve the problem of multiple reads of the buffer for a single dump then a) that is not a problem on physical hardware and b) it does not actually solve that problem unless you take a snapshot on file open and release the snapshot on file close.


Which means your 16MB buffer now requires 4096 separate
reads! And you only doing partial reads of the section you think you
need is never going to be reliable on live system. Not sure why you
would want to anyway. It is just making things much more complex. You
now need an intelligent user land program to read the log out and decode
I don't need it. We can add it later. And we can add it on top what we
already expose without the need to recompile/rebuild the driver.
Or we could put that effort into something which will be of significant benefit to all users of the interface.


at least the header section to know what data section to read. You can't
just dump the whole thing with 'cat' or 'dd'.
only 'cat' wont work as it's binary file
Which is still a problem.


for reliable output via dmesg - see my proposal at [1]

[1] https://patchwork.freedesktop.org/series/133613/
Hmm. Actually, is this interface allowing the filesystem layers to issue
multiple read calls to read the buffer out in small chunks? That is also
going to break things. If the GuC is still writing to the log as the
user is reading from it, there is the opportunity for each chunk to not
follow on from the previous chunk because the data has just been
overwritten. This is already a problem at the moment that causes issues
when decoding the logs, even with an almost atomic copy of the log into
a temporary buffer before reading it out. Doing the read in separate
chunks is only going to make that problem even worse.
current solution, that converts data into hex numbers, reads log buffer
in chunks of 128 dwords, how proposed here solution that reads in 4K
chunks could be "even worse" ?
See above, 4KB chunks means 4096 separate reads for a 16M buffer. And
each one of those reads is a full round trip to user land and back. If
but is this a proven problem for us?
So what problem are you trying to solve?


you want to get at all close to an atomic read of the log then it needs
to be done as a single call that copies the log into a locally allocated
kernel buffer and then allows user land to read out from that buffer
rather than from the live log. Which can be trivially done with the
current method (at the expense of a large memory allocation) but would
be much more difficult with random access reader like this as you would
need to say the copied buffer around until the reads have all been done.
Which would presumably mean adding open/close handlers to allocate and
free that memory.
as I mentioned above if we desperately need larger copies then we can
use the code promoted from the SOUND subsystem
Even more effort.


but for random access reader (up to 4K) this is what this patch already
provides.
But why do you need random access? Streaming? Then implement LFDs.



and in case of some smart tool, that would understands the layout of the
GuC log buffer, we can even fully eliminate problem of reading stale
data, so why not to choose a more scalable solution ?
You cannot eliminate the problem of stale data. You read the header, you
read the data it was pointing to, you re-read the header and find that
the GuC has moved on. That is an infinite loop of continuously updating
pointers.
I didn't say that I can create snapshot that is 100% free of stale data,
what I meant was that with this proposal I can provide almost real time
access to the GuC log, so with custom tool I can read pointers and and
log entries as small randomly located chunks in the buffer, without the
need to output whole log buffer snapshot as giant text file that I would
have to parse again.
But why do you want to read small chunks?

John.


John.

John.

Signed-off-by: Michal Wajdeczko <michal.wajdeczko@xxxxxxxxx>
Cc: Lucas De Marchi <lucas.demarchi@xxxxxxxxx>
Cc: John Harrison <John.C.Harrison@xxxxxxxxx>
---
Cc: linux-fsdevel@xxxxxxxxxxxxxxx
Cc: dri-devel@xxxxxxxxxxxxxxxxxxxxx
---
    drivers/gpu/drm/xe/xe_guc_debugfs.c | 26 ++++++++++++++++++++++++++
    1 file changed, 26 insertions(+)

diff --git a/drivers/gpu/drm/xe/xe_guc_debugfs.c
b/drivers/gpu/drm/xe/xe_guc_debugfs.c
index d3822cbea273..53fea952344d 100644
--- a/drivers/gpu/drm/xe/xe_guc_debugfs.c
+++ b/drivers/gpu/drm/xe/xe_guc_debugfs.c
@@ -8,6 +8,7 @@
    #include <drm/drm_debugfs.h>
    #include <drm/drm_managed.h>
    +#include "xe_bo.h"
    #include "xe_device.h"
    #include "xe_gt.h"
    #include "xe_guc.h"
@@ -52,6 +53,29 @@ static const struct drm_info_list debugfs_list[] = {
        {"guc_log", guc_log, 0},
    };
    +static ssize_t guc_log_read(struct file *file, char __user *buf,
size_t count, loff_t *pos)
+{
+    struct dentry *dent = file_dentry(file);
+    struct dentry *uc_dent = dent->d_parent;
+    struct dentry *gt_dent = uc_dent->d_parent;
+    struct xe_gt *gt = gt_dent->d_inode->i_private;
+    struct xe_guc_log *log = &gt->uc.guc.log;
+    struct xe_device *xe = gt_to_xe(gt);
+    ssize_t ret;
+
+    xe_pm_runtime_get(xe);
+    ret = xe_map_read_from(xe, buf, count, pos, &log->bo->vmap,
log->bo->size);
+    xe_pm_runtime_put(xe);
+
+    return ret;
+}
+
+static const struct file_operations guc_log_ops = {
+    .owner        = THIS_MODULE,
+    .read        = guc_log_read,
+    .llseek        = default_llseek,
+};
+
    void xe_guc_debugfs_register(struct xe_guc *guc, struct dentry
*parent)
    {
        struct drm_minor *minor = guc_to_xe(guc)->drm.primary;
@@ -72,4 +96,6 @@ void xe_guc_debugfs_register(struct xe_guc *guc,
struct dentry *parent)
        drm_debugfs_create_files(local,
                     ARRAY_SIZE(debugfs_list),
                     parent, minor);
+
+    debugfs_create_file("guc_log_raw", 0600, parent, NULL,
&guc_log_ops);
    }





[Index of Archives]     [Linux Ext4 Filesystem]     [Union Filesystem]     [Filesystem Testing]     [Ceph Users]     [Ecryptfs]     [NTFS 3]     [AutoFS]     [Kernel Newbies]     [Share Photos]     [Security]     [Netfilter]     [Bugtraq]     [Yosemite News]     [MIPS Linux]     [ARM Linux]     [Linux Security]     [Linux Cachefs]     [Reiser Filesystem]     [Linux RAID]     [NTFS 3]     [Samba]     [Device Mapper]     [CEPH Development]

  Powered by Linux