OpenGL RGB DXT1 compressed texture mipmap upload

I'm trying to upload GL_COMPRESSED_RGB_S3TC_DXT1_EXT format mipmaped texture levels using PBO. the program is resizing and compressing the images using the stb_resize.h and stb_dxt.h libraries.

Apparently the image compression works okay, but uploading any mipmap levels produces following result:

However If I define no mipmap levels for the texture it renders correcly. Also I'm not getting any errors from OpenGL as I have GL_KHR_debug enabled.

Here is the opengl code for uploading mipmaps, with some of my own library code in between:

// Load image:
auto img = loadImage2D("parrot.png");

// Resize the image to power-of-two
int max_mipmaps = 1;
unsigned int max_size = nextPow2(std::max(img.width, img.height));
for(int i = max_size; i > 4; i >>= 1)
    max_mipmaps++;

auto rimg = img.scaleCubic(max_size,max_size);

// Allocate PBO to source compressed mipmap levels from
handle_t pbo;
size_t pbosize = rimg.compressed_size();
glCreateBuffers(1, &pbo);
glBindBuffer(GL_PIXEL_UNPACK_BUFFER, pbo);
glBufferStorage(GL_PIXEL_UNPACK_BUFFER, pbosize, 0,
    GL_DYNAMIC_STORAGE_BIT|GL_MAP_WRITE_BIT|GL_MAP_READ_BIT|GL_MAP_PERSISTENT_BIT);
// Map the buffer
void * pbodata = (char*)glMapBufferRange(
    GL_PIXEL_UNPACK_BUFFER, 0, pbosize,
    GL_MAP_WRITE_BIT|GL_MAP_READ_BIT
    |GL_MAP_PERSISTENT_BIT
    |GL_MAP_FLUSH_EXPLICIT_BIT);

// Compress the image and write the data directly into PBO.
// compress_to() fmt returns GL_COMPRESSED_RGB_S3TC_DXT1_EXT
option_t fmt;
rimg.compress_to(pbodata, fmt);
glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, 0, pbosize);

// Allocate the texture
glCreateTextures(GL_TEXTURE_2D, 1, &texture);
glBindTexture(GL_TEXTURE_2D, texture);
glTextureStorage2D(texture, max_mipmaps, fmt, rimg.width, rimg.width);
// Upload base image level.
glPixelStorei(GL_UNPACK_ALIGNMENT, 1);
glCompressedTextureSubImage2D(texture, 0, 0,0, rimg.width, rimg.height, fmt, rimg.compressed_size(), 0);
// Process and Upload mipmap levels.
unsigned int mipmapsize = max_size >> 1;
for(int i = 1; i < max_mipmaps && mipmapsize >= 4; ++i) {
    // Resize next mipmap level.
    rimg = img.scaleLinear(mipmapsize,mipmapsize);

    // Compress the image and write result to *pbodata
    rimg.compress_to(pbodata, fmt);
    glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, 0, pbosize);

    // Upload mipmap image level.
    glCompressedTexSubImage2D(GL_TEXTURE_2D, i, 0,0, rimg.width, rimg.height, fmt, rimg.compressed_size(), 0);

    mipmapsize >>= 1;
}
// Discard the PBO
glUnmapNamedBuffer(pbo);
glBindBuffer(GL_PIXEL_UNPACK_BUFFER,0);
glDeleteBuffers(1, &pbo);

// Set texture params.
glTextureParameteri(texture, GL_TEXTURE_BASE_LEVEL, 0);
glTextureParameteri(texture, GL_TEXTURE_MAX_LEVEL, max_mipmaps - 1);
glTextureParameteri(texture, GL_TEXTURE_MIN_FILTER, GL_LINEAR_MIPMAP_LINEAR);
glTextureParameteri(texture, GL_TEXTURE_MAG_FILTER, GL_LINEAR)

Doing glFlushMappedBufferRange will not be sufficient to ensure proper synchronization with the GPU. And this is what seems to be happening here: the whole mipmap pyramid is created on the CPU, before the GPU processes the first glCompressedTexSubImage2D call.

As a quick&dirty test for this hypothesis, you can add a glFinish into the loop. To properly synchronize it, you could use GL Sync Objects, but it would be best if you avoided any need to synchronize at all: just use a PBO which is big enough to hold the whole mipmap pyramid data.