Hi Christian, I agree with the crux of your argument but you can crash the system with carefully timed read/writes from the GFX registers, or MC, or ... That's why the entries are root only. The issue with locking excessively is you might change behaviour of something you're trying to debug because the kernel then has to wait on the lock (or you're waiting on the lock). For instance, suppose you want to read DCE registers and the kernel is doing some gfx/grbm related your DCE read will either block the GFX activity or you're waiting on it which means the behaviour of the system could be different than from the case where you're not probing the registers. The goal is to be as unobtrusive as possible. Heck I often mmap the BAR so I can read MMIO registers directly and not waste CPU cycles on syscalls. The PM lock is because reading PGFSM registers during a powerup/down results in a GPU hang, and we only take the GRBM lock when users want to read from specific banks. For any other case we don't need a lock. In the kernel they should be mutually exclusive. The PM lock is only taken in PP when performing a transition (or initializing the core) and I have yet to see a place where PP/PM related code also grabs the grbm index lock. I don't have strong feelings against your proposal but I do prefer not to take the locks when I don't need to. Ultimately I'll leave the decision up to you and Alex. I'd like to get the 2nd patch through though as it makes the write side of the mmio reg complimentary (e.g. bank selection/pm lock). Cheers, Tom ________________________________ From: Christian König <deathsimple@xxxxxxxxxxx> Sent: Tuesday, October 11, 2016 08:11 To: StDenis, Tom; amd-gfx at lists.freedesktop.org Subject: Re: [PATCH] drm/amd/amdgpu: Make debugfs write compliment read The only downside to taking the locks all the time is for 99% of debug reads it's unnecessary and in theory could lead to performance/behaviour problems. Yeah, but performance isn't critical here and taking two locks compared to the rest of what we (IOCTL etc...) do is completely negligibly. I've been using this on read for a while now and haven't noticed a problem. My concern is rather that somebody tries to write to the PM regs on accident or purpose without taking the lock. See, the kernel is usually responsible to keep people from doing stupid things with the hardware, even if those people are root and could do stupid things anyway. Well, you know what I mean? Regards, Christian. Am 10.10.2016 um 22:21 schrieb StDenis, Tom: The only downside to taking the locks all the time is for 99% of debug reads it's unnecessary and in theory could lead to performance/behaviour problems. I've been using this on read for a while now and haven't noticed a problem. Tom ________________________________ From: Christian König <deathsimple@xxxxxxxxxxx><mailto:deathsimple at vodafone.de> Sent: Monday, October 10, 2016 09:49 To: Tom St Denis; amd-gfx at lists.freedesktop.org<mailto:amd-gfx at lists.freedesktop.org> Cc: StDenis, Tom Subject: Re: [PATCH] drm/amd/amdgpu: Make debugfs write compliment read Would it hurt us to always take the looks (both the pm as well as grbm idx lock)? I don't think that would be much of an issue and I would have a better feeling with that. Christian. Am 10.10.2016 um 15:38 schrieb Tom St Denis: that's kind of the point. Its for debugging only. Also reads can already lock On Mon, Oct 10, 2016, 09:25 Christian König <deathsimple at vodafone.de<mailto:deathsimple at vodafone.de>> wrote: Mhm, that userspace controls if the lock is taken or not is a bit problematic, on the other hand only root could access that. Christian. Am 10.10.2016 um 14:51 schrieb Tom St Denis: > Add PG lock support as well as bank selection to > the MMIO write function. > > Signed-off-by: Tom St Denis <tom.stdenis at amd.com<mailto:tom.stdenis at amd.com>> > --- > drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 43 ++++++++++++++++++++++++++++++ > 1 file changed, 43 insertions(+) > > diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c > index 04c4aee70452..a9d83c1d7a43 100644 > --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c > +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c > @@ -2637,10 +2637,45 @@ static ssize_t amdgpu_debugfs_regs_write(struct file *f, const char __user *buf, > struct amdgpu_device *adev = f->f_inode->i_private; > ssize_t result = 0; > int r; > + bool pm_pg_lock, use_bank; > + unsigned instance_bank, sh_bank, se_bank; > > if (size & 0x3 || *pos & 0x3) > return -EINVAL; > > + /* are we reading registers for which a PG lock is necessary? */ > + pm_pg_lock = (*pos >> 23) & 1; > + > + if (*pos & (1ULL << 62)) { > + se_bank = (*pos >> 24) & 0x3FF; > + sh_bank = (*pos >> 34) & 0x3FF; > + instance_bank = (*pos >> 44) & 0x3FF; > + > + if (se_bank == 0x3FF) > + se_bank = 0xFFFFFFFF; > + if (sh_bank == 0x3FF) > + sh_bank = 0xFFFFFFFF; > + if (instance_bank == 0x3FF) > + instance_bank = 0xFFFFFFFF; > + use_bank = 1; > + } else { > + use_bank = 0; > + } > + > + *pos &= 0x3FFFF; > + > + if (use_bank) { > + if ((sh_bank != 0xFFFFFFFF && sh_bank >= adev->gfx.config.max_sh_per_se) || > + (se_bank != 0xFFFFFFFF && se_bank >= adev->gfx.config.max_shader_engines)) > + return -EINVAL; > + mutex_lock(&adev->grbm_idx_mutex); > + amdgpu_gfx_select_se_sh(adev, se_bank, > + sh_bank, instance_bank); > + } > + > + if (pm_pg_lock) > + mutex_lock(&adev->pm.mutex); > + > while (size) { > uint32_t value; > > @@ -2659,6 +2694,14 @@ static ssize_t amdgpu_debugfs_regs_write(struct file *f, const char __user *buf, > size -= 4; > } > > + if (use_bank) { > + amdgpu_gfx_select_se_sh(adev, 0xffffffff, 0xffffffff, 0xffffffff); > + mutex_unlock(&adev->grbm_idx_mutex); > + } > + > + if (pm_pg_lock) > + mutex_unlock(&adev->pm.mutex); > + > return result; > } > -------------- next part -------------- An HTML attachment was scrubbed... URL: <https://lists.freedesktop.org/archives/amd-gfx/attachments/20161011/5797a32a/attachment-0001.html>
